Information Fusion 100 (2023) 101906

Contents lists available at ScienceDirect

Information Fusion
journal homepage: www.elsevier.com/locate/inffus

A survey on fairness-aware recommender systems✩

Di Jin a , Luzhi Wang a , He Zhang b , Yizhen Zheng b , Weiping Ding c ,∗, Feng Xia d , Shirui Pan e ,∗
a
College of Intelligence and Computing, Tianjin University, China
b
Department of Data Science and AI, Faculty of IT, Monash University, Australia
c
School of Information Science and Technology, Nantong University, China
d
School of Computing Technologies, RMIT University, Australia
e School of Information and Communication Technology, Griffith University, Australia

ARTICLE INFO ABSTRACT

Keywords: As information filtering services, recommender systems have extremely enriched our daily life by providing
Recommender systems personalized suggestions and facilitating people in decision-making, which makes them vital and indispensable
Fairness to human society in the information era. However, as people become more dependent on them, recent
Trustworthiness
studies show that recommender systems potentially own unintentional impacts on society and individuals
Survey
because of their unfairness (e.g., gender discrimination in job recommendations). To develop trustworthy
services, it is crucial to devise fairness-aware recommender systems that can mitigate these bias issues. In
this survey, we summarize existing methodologies and practices of fairness in recommender systems. Firstly,
we present concepts of fairness in different recommendation scenarios, comprehensively categorize current
advances, and introduce typical methods to promote fairness in different stages of recommender systems.
Next, after introducing datasets and evaluation metrics applied to assess the fairness of recommender systems,
we will delve into the significant influence that fairness-aware recommender systems exert on real-world
industrial applications. Subsequently, we highlight the connection between fairness and other principles of
trustworthy recommender systems, aiming to consider trustworthiness principles holistically while advocating
for fairness. Finally, we summarize this review, spotlighting promising opportunities in comprehending
concepts, frameworks, the balance between accuracy and fairness, and the ties with trustworthiness, with
the ultimate goal of fostering the development of fairness-aware recommender systems.

1. Introduction their attributes information such as annual income, address, occupa-

tion, and so on [12]. Lastly, in the education industry, Massive Open
Recommendation systems (RSs) are information filtering systems Online Course (MOOC), one of the leading online learning providers,
that are expected to suggest products and services, i.e., items, that recommends courses to users based on their historical opinions [13].
most likely interest a user [1]. The suggestions are related to vari- However, as the development of recommender systems surges and
ous decision-making processes for a user, such as which products to the reliance on them grows, these systems can lead to huge negative
purchase, which videos to watch and which songs to listen. As the impacts on society and individuals due to unfairness. The causes of un-
world becomes more information-overloaded, recommender systems fairness and their undesirable outcomes are numerous and significant.
are particularly useful when users need to choose an item from an For instance, a loan domain recommender system may recommend
overwhelming selection offered by a service. Recommender systems loan domains influenced by a user’s sensitive personal attributes, such
are pervasive and have been utilized in various fields including e-
as age, gender, and race. As a result, minority groups such as senior
commerce [2–4], economics [5–7], education [8–10], etc. For the
citizens, women workers, and ethnic minorities may suffer worsening
e-commerce industry, Amazon, for example, distributes product infor-
financial situations by being recommended loan domains with higher
mation provided by merchants to users based on their history of pur-
interest rates, which is obviously unfair. The second example is MOOC
chases or website interaction. Remarkably, its recommendation engine
course recommendations [13], in which more than 40% of the courses
accounts for three-quarters of its total revenue [11]. In the economics
are taught by American teachers, while the remaining courses from 73
industry, the loans domain can be recommended to users based on

✩ This project is funded by the National Natural Science Foundation of China (No: 62272340 and 61976120) and ARC Future Fellowship (No. FT210100097).
∗ Corresponding authors.
E-mail addresses: [email protected] (D. Jin), [email protected] (L. Wang), [email protected] (H. Zhang), [email protected] (Y. Zheng),
[email protected] (W. Ding), [email protected] (F. Xia), [email protected] (S. Pan).

https://doi.org/10.1016/j.inffus.2023.101906
Received 29 May 2023; Received in revised form 27 June 2023; Accepted 28 June 2023
Available online 4 July 2023
1566-2535/© 2023 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
D. Jin et al. Information Fusion 100 (2023) 101906

Fig. 1. The organizational layout of this survey.

countries are rarely recommended to or enrolled in by online users. Due fairness) or dynamic evolution (e.g., dynamic fairness [27]) should be
to the vast geographic imbalance in MOOC recommendations, teachers taken into account since recommender systems interact with both users
in smaller or less-known places faced a great disadvantage when at- and item providers in a dynamic process. Therefore, achieving fairness
tracting students. In a similar way, Amazon tended to recommend items for different recommender systems remains a complex task. (2) Full
from larger merchants over smaller merchants. Because of this, smaller life cycle of fairness demands. Unfairness exists in the whole life-cycle
merchants would find it difficult to compete with large merchants, of recommender systems because of their dynamic interactions with
even if they offered better prices or quality. Although recommender users and items (as shown in Fig. 2). This fact requires researchers and
systems are supposed to mitigate these unfairnesses, they can actually practitioners to enhance fairness at each stage of developing recom-
exacerbate them in the recommendation pipeline. mender systems to avoid learning historical unfairness from the last
Driven by these unintentional issues, people increasingly yearn for loop and even amplifying it into the subsequent feedback loop. (3)
fairness since it is critical and essential in developing recommender Adverse effects of improving fairness. Existing studies have demonstrated
systems that can be trusted. First, fairness benefits users, items, and the existence of the trade-off between fairness and accuracy in recom-
even recommender systems themselves [14]. For example, in a fair mender systems [28], which makes designing recommender systems
recommender system, users can obtain more relevant information, that simultaneously prioritize fairness while maintaining satisfactory
including niche information, which can aid in breaking out of the performance not easy. Moreover, improving fairness of recommender
cocoon of information. As minor items are allocated more exposure, systems can potentially influence other aspects (e.g., robustness, pri-
the Matthew effect [15] is lessened, encouraging providers to enhance vacy) of trustworthiness. For example, a recent work shows that in-
their creativity and diversity. By providing equal quality of service dividual fairness is at the cost of privacy [29]. Thus, evaluating the
for objects from diverse backgrounds, fair recommender systems can influence of fairness should be taken into account when comprehen-
also gain long-term interest due to positive feedback from users and sively building fairness-aware recommender systems. Therefore, it is
item providers. Second, a global consensus has recently been built on imperative to summarize current efforts and advancements in building
enhancing the trustworthiness of AI systems [16–18], including fair- fairness-aware recommender systems, which contribute to developing
ness, robustness, explainability, privacy, etc. Devising fairness-aware trustworthy, responsible, and socially beneficial AI services.
recommender systems directly contribute to the trustworthiness of In this survey, fair recommender systems refer to recommender
RSs [18]. Moreover, studying the connections between fairness and systems that can adapt to different users/items and provide indiscrimi-
other aspects (e.g., robustness [19]) benefits the comprehensive build- nate recommendation services to them. Related reviews include surveys
ing of trustworthy systems [16,17]. Finally, compliance with laws on recommender systems, fair recommender systems, and trustworthy
and regulations [14] requires recommender systems to be fair when recommender systems. Our survey differs from those surveys in that it
interacting with people because fairness is one of the cornerstones of elaborates on the existing advancements in the fair recommender sys-
keeping social order. For example, discrimination against vulnerable tems as well as how fairness interacts with other aspects of trustworthy
groups of people based on sensitive information (e.g., gender, age, recommender systems. (1) Surveys on recommender systems review
race [20,21]) is forbidden by current anti-discrimination laws [22], the advancements of performance-oriented methods on recommender
which also require that similar people should be treated similarly to systems. Wu et al. [30] explore the impact of Graph Neural Networks on
ensure equality. different categories of recommender systems. Deldjoo et al. [31] inves-
Although the fairness of general machine learning tasks (e.g., image tigate the impact of adversarial learning on the security and accuracy of
classification) has been extensively explored [23], building fairness- recommender systems. Wu et al. [32] introduce the modeling of collab-
aware recommender systems is no-trivial since the following chal- orative filtering techniques in different types of recommender systems
lenges. (1) Diverse and unique fairness concepts. Various real-world (e.g., content-rich recommender systems, sequential recommender sys-
scenarios of RSs encompass diverse fairness concepts, such as group or tems). (2) Recently, surveys on fair recommendation systems have also
individual fairness [24,25], static or dynamic fairness [26,27]. More- emerged. For example, Wang et al. [14] explore various perspectives
over, the unique characteristics (e.g., severing sellers and buyers si- on unfairness issues and provide an overview of existing approaches,
multaneously, and changing recommendation trends driven by time) which include data-oriented, ranking, and re-ranking methods. Zehlike
of recommender systems call for elaborated fairness concepts. For et al. [33] delve into how ranking methods impact the fairness of
example, the fairness concerning multi-party benefits (e.g., multi-sided recommender systems. They introduce the fair ranking framework and

2
D. Jin et al. Information Fusion 100 (2023) 101906

Fig. 2. The lifecycle of recommender systems. In the lifecycle, the data collection phase collects data from users; the model learning phase feeds data into the model for training;
the loop phase provides recommendation results to users. In the lifecycle, biases in each phase will have an unfair impact on the recommender system. With the operation and
feedback loop of recommender systems, existing biases can be potentially amplified in the following phases of recommender systems.

summarize the evaluation of fair ranking methods. Chen et al. [34] • Building Connections between Fairness and Other Ethical
describe the existence of biases in recommender systems and describe Principles. This study demonstrates how fairness relates to other
ways to address them. Unlike these surveys, our survey categorizes ethical principles in trustworthy recommendation systems. It en-
current methods into pre-processing, in-processing, and post-processing courages people to consider holistically while advocating fairness.
methods; we also present comprehensive and fine-grained taxonomy A question such as, does promoting fairness affect other ethical di-
on each of them (e.g., data re-labeling, data re-sampling, and data mensions of trustworthy recommendation systems, e.g., explain-
modification in pre-processing methods). (3) Current surveys on trust- ability and robustness, should be asked.
worthy recommender systems allocate their attention to several dif- • Evaluation of Existing Challenges and Future Direction. We
ferent aspects of trustworthiness (e.g., explainability, robustness) [18, highlight the existing limitations and challenges of existing
35] or present a conceptual framework of trustworthy recommender fairness-promoting methods for recommender systems. In future
systems [36]. In contrast, our survey focuses on comprehensively sum- works, these problems should be further considered and ad-
marizing current advancements in fairness and discussing the influence dressed. In particular, the definition of fairness is not consistent
on other aspects of trustworthy recommender systems from the view of between studies, which can easily cause confusion. Additionally,
enhancing fairness. despite improvements in fairness, existing methods neglect the
relationship between fairness and other ethical dimensions of
We introduce the theory and practice of fairness-aware recom-
trustworthy systems, which can deteriorate other ethical metrics.
mender systems in this survey. As a first step, we describe the concepts
and formulations of recommender systems and fairness in Section 2.
2. Concepts and formulations
Then, we categorize and illustrate methods promoting fairness in dif-
ferent stages of recommender systems in Section 3 after carefully
2.1. Recommender systems
analyzing related literature regarding fairness in recommender sys-
tems. Specifically, these processing stages include pre-processing, in-
A recommender system serves as an information filtering system
processing, and post-processing. In Section 4, we collect and collate that learns and predicts the user’s interest in an item [37]. According to
datasets and evaluation metrics used in literature exploring fairness relevant information such as items, users, and the interaction between
in recommender systems. Following that, we discuss the industrial them, the recommender system provides users with personalized ser-
applications of recommendation systems that consider fairness in e- vices and recommends suitable items to them [38]. Given a user 𝑢 ∈ 𝑈 ,
commerce, education, and social activities in Section 5. Moreover, an item 𝑣 ∈ 𝑉 , where 𝑈 is a user set and 𝑉 is an item set, the goal of
Section 6 explores the connection between fairness and other ethical the recommender system is to learn an information filter 𝑓 (⋅) to capture
principles of a trustworthy recommender system such as explainability, user preference. The predicted score 𝑦𝑢,𝑣 for the user’s preferences in
robustness, privacy, and so on. Section 7 provides a big picture of items is [30]:
future directions of fairness-aware recommender systems from different
dimensions, including concepts, frameworks, trade-off, and trustworthi- 𝑦𝑢,𝑣 = 𝑓 (𝑢, 𝑣). (1)
ness. Last but not least, Section 8 summarizes this survey’s importance The lifecycle of the recommender system can be abstracted into a
and influence within the context of trustworthy recommender systems. feedback loop composed of users, data, and models. The loop consists
Fig. 1 illustrates the organizational layout of this survey, as well as of three phases as shown in Fig. 2, including, a data collection phase,
the logical connections between each section. The contributions of this a model learning phase, and a feedback phase. The data collection
survey can be enumerated as follows: phase is established between users and data, and it collects users’
and items’ attribute information, as well as user–item interaction data.
• A Holistic Taxonomy of Fairness-aware RS Methods. Com- The learning phase is built with the data and the model. In specific,
pared to previous surveys, this study presents a more comprehen- this refers to the development of a recommendation model based on
sive taxonomy for methods that improve fairness in recommender the collected data. A recommender system utilizes historical data to
systems. These methods are grouped in accordance with their forecast the probability of an item being recommended to a user. This
roles in the three phases of implementing recommender systems: step delivers the suggested results to users to satisfy their information
pre-processing, in-processing, and post-processing. A useful and needs. This phase will impact users’ future behavior and decisions.
intuitive framework is provided for the adoption or understanding The recommender system can be divided into several types accord-
of these methods by interested researchers. ing to the different scenarios:

3
D. Jin et al. Information Fusion 100 (2023) 101906

Fig. 3. Fairness-aware methods in the three stages. The pre-processing stage encompasses several methods like data re-labeling, data re-sampling, and data modification. These
pre-processing methods are independent of recommendation models. The in-processing stage is mainly to improve the fairness of the model from a method perspective. The
in-processing stage involves various techniques, including regularization-based methods, causal-inference-based methods, adversarial-learning-based methods, reinforcement-learning-
based methods, and ranking-optimization-based methods. The post-processing stage also employs model-independent methods, treating the model as a black box and solely processing
the model’s output results. The post-processing methods include non-parametric re-ranking methods and parametric re-ranking methods.

• Session-based RSs take each session as the input unit, captur- recommendations, which may cause undesirable or even disastrous
ing the user’s short-term preferences and dynamic interests as consequences for human life and society. Next, we will introduce some
reflected in session transitions [39]. common biases in datasets.
• Conversational-based RSs are recommender systems that can User Bias. User attribute bias is a common user bias. In recom-
interact with users in multiple rounds in real-time, eliciting users’ mender systems, some sensitive attributes like age [44], geographical
dynamic preferences and taking actions based on their current location [45], gender [46], and profession [47] are important sources of
needs [40]. information for recommender systems to understand user preferences.
• Content-enriched RSs incorporate some auxiliary data related to However, sometimes users’ attributes (e.g., age, gender) can cause
users and items to enhance representation learning and seman- biased recommendation results. For example, age may be an effective
tic relevance. These auxiliary data may include textual content, feature in a music recommender system as they can recommend music
knowledge graphs, etc [32]. to users according to their ages [48]. In general, youngsters have
• Social RSs are defined as any recommender system for the social a higher preference for hip-hop music than senior people. However,
media domain. This type of recommender systems improves rec- sometimes users of a certain age group may wish to jump out of the
ommendation performance by incorporating social relations into cocoon of age information and explore different kinds of music. For
the recommender system [13]. the gender attribute, some researchers [49–51] observe the promotion
of high-paying jobs varies by gender. In MOOC, a recommender system
There are unique unfairnesses in different types of recommender sys- attempt to guide course resources to students [52], the geographic
tems, and we introduce these unfairness issues and the biases that location and nationality of teachers also influence which courses are
generate unfairness in the follow-up. recommended to students [13].
In addition to user attribute bias, user selection bias is also an
2.2. Fairness important part of user bias. User selection bias refers to the behavioral
bias when the user selects items, which usually exists in user explicit
Fairness is a concept that originated in sociology, economics, and feedback. The existence of biases in user feedback data can lead to
law [41–43]. Its definition in the Oxford English Dictionary is ‘‘imper- inconsistencies between user preferences and behavior records [53].
fect and just treatment or behavior without favoritism or discrimina- For instance, music streaming media recommender systems provide
tion’’. In the context of recommender systems, fairness requires that playlists based on the music preference of users [54], which can be
recommender systems treat all users and items equally. For example, in affected by the user feedback bias. Specifically, during the recommen-
loan approval based on recommender systems [12], the systems are fair dation process, the model is updated online based on user feedback.
if the approval of an applicant seeking a loan is not influenced by the In a music recommender system, explicit feedback (e.g., ratings of
user’s attributes (e.g., gender, race). We can develop fair recommender items [55]) can be the track marked ‘‘favorite’’ by the user. However,
systems when bias or unfairness is eliminated from systems [14]. In users may have wrongly clicked like tags, which will lead to explicit
this section, we first categorize causes of unfairness and then present feedback bias [56].
common fairness expressions that can measure the existence of bias. Exposure Bias. Exposure Bias refers to the situation where users
only have access to a portion of the available item set [57]. The
2.2.1. Causes of unfairness exposure bias is often present in implicit feedback data, which usually
As depicted in Fig. 2, the unfairness of a recommender system refers to whether the user interacts with the item, including some
appears in its whole lifecycle, including the data collection phase, the purchases, clicks, and other behaviors [58]. For example, e-commerce
model learning phase, and the feedback loop phase. According to the sites like Amazon allow users to provide feedback on recommended
position of bias in the three phase, we divide bias in the recommender items beyond the user interface by searching and browsing various
system into three categories, namely the data bias, model bias, and product pages. The user’s browsing behaviors may contain incorrect
feedback bias. Subsequently, we will elucidate these biases by showing clicks, and these mistakes can make the recommender system misjudge
how they result in unfairness in recommender systems. the user’s preference, which results in an unfair recommendation [11].
Data Bias. Currently, recommender systems are always trained In a music recommender system, a handful of popular artists may
on large datasets, which usually contain user information like user garner the majority of traffic, thereby underexposing less mainstream
behaviors (e.g. incorrect clicks), sensitive attributes (e.g. gender), and artists. If this bias remains unaddressed, recommender systems could
the interaction between users and items. However, training datasets adversely affect the experience of diverse users and items on the
potentially include various biases. These biases can lead to unfair platform due to continuous interaction with biased recommendations,

4
D. Jin et al. Information Fusion 100 (2023) 101906

and thus the training of models using biased interactions in subsequent the updated data for self-reinforcing. This feedback loop mechanism not
timeframes [26]. In music streaming media recommender systems, the only generates bias, but also exacerbates the bias over time, resulting in
explicit feedback can be the track marked ‘‘favorite’’ by the user, and a gradual deterioration of the fairness ecosystem. In the following para-
implicit feedback can be the number of times a track has been played. graph, we introduce the popularity bias, which is a typical feedback
However, users may pay less attention to the music app while doing bias.
other activities, such as exercising, reading late at night, or commuting, Popularity Bias. A few popular items are frequently recommended,
and the music is looped multiple times, which can lead to an invisible while most others are disregarded. Users consume these recommenda-
feedback bias [56]. tions, and their responses are recorded and added to the system. Over
Time Bias. Some time-related recommendations, such as news rec- time, the recommender system recommends popular items to users,
ommendations, session-based recommendations, and job recommen- continuously collects user feedback on popular items, and adds them
dations, heavily rely on direct user–item interactions to understand to the training set, making the data distribution more unbalanced,
user preferences and provide specialized recommendations. However, which will result in more and more recommendation results focusing on
freshly released data could unfairly overrepresent users’ long-term popular items [65]. The existence of popularity bias will also constantly
interests, which is unfair to learn users’ preferences. If we continue to change the user’s preference representation, making it challenging for
disclose items at time 𝑡+𝑛 according to the fairness limitations at time 𝑡, the recommender system to capture the user’s true preference. For
even though an item that is popular at time 𝑡 may no longer be popular example, Naghiaei et al. [66] investigate the impact of popularity issues
at time 𝑡 + 𝑛, we will neglect the long-term fairness dynamics [27]. on book recommender systems. Their work shows that recommender
This will ignore the long-term dynamic process of fairness, which leads systems tend to recommend popular items frequently, and there is a
to recommended bias. For example, the goal of job recommendations strong correlation existing between the popularity of books and the
is to recommend job advertisements to job seekers. Job seekers prefer frequency with which they are recommended. Most books are not
to click after seeing a new job advertisement. New job advertisements exposed to users by the recommender system, while popular books are
obtain higher click rates as a result of this behavior than older job highlighted more frequently.
advertisements. Consequently, jobs of longer-lasting professionals are
less likely to be recommended. However, job seekers are driven by their 2.2.2. Fairness expressions in recommender systems
long-term career aspirations when contemplating jobs. Advertising rec- A recommender system can be employed in multiple scenarios, and
ommendations that are consistent with these preferences are more the unfairness concerns in each scenario are different. Next, we present
advantageous to job seekers [59]. Therefore, it is unfair to recommend the expressions of fairness in recommender systems from different
recent advertisements to users without considering users’ long-term perspectives.
interests. Individual Fairness vs. Group Fairness. Given two similar users
Cold-start Bias. The above biases only consider unfairness in the case 𝑢𝑖 , 𝑢𝑗 ∈ 𝑈 , individually fairness-aware recommender systems hope to
of warm-start recommendation. The primary source of unfairness in this give similar prediction scores for samples 𝑢𝑖 and 𝑢𝑖 [24]. For exam-
instance is data biases (e.g., clicks or pageviews). However, the recom- ple, in a healthcare recommendation system, two patients exhibiting
mender system will meet a cold start problem when lacks data. When similar pathologies should receive recommendations of equivalent qual-
a new user or item enters the system, the cold-start problem occurs ity [67]. Methods for individual fairness solve the problem of statistical
at which point the recommender system fails to provide personalized equality through pairwise comparisons between similar users. The
recommendations because it lacks sufficient data on user behaviors formal definition of individual fairness can be expressed as 𝑓 (𝑢𝑖 , 𝑣) ≈
or item attributes [60]. Solving a cold-start problem is usually using 𝑓 (𝑢𝑗 , 𝑣). where 𝑓 (⋅) is a recommender system, 𝑣 is an item.
prior knowledge from warm-start recommendations to train cold-start Group fairness necessitates that protected groups receive treatment
recommender systems. The cold-start bias refers to the bias of this prior akin to that of advantaged groups [25]. As a typical group fairness,
knowledge in warm-start recommender systems. These biases will be demographic parity [15] requires that the probability that the pro-
brought into the cold-start recommender system when training, and tected group (e.g., female group) is predicted to be a positive sam-
cause the unfairness phenomena [61]. The unfairness phenomena can ple (e.g., job offers) is equal to the probability that the advantaged
be particularly problematic because the unfairness caused by cold-start group (e.g., male group) is predicted to be a positive sample. Given
recommendations can persist and accumulate throughout the lifespan a protected group 𝑈𝑖 and an advantaged group 𝑈𝑗 , the demographic
of the item, making it increasingly difficult to mitigate unfairness [62]. parity of the recommender system can be formalized as 𝑃 𝑟(𝑦𝑈𝑖 ,𝑣 =
Model Bias. The model learning phase uses the collected data in 1|𝑈𝑖 ) = 𝑃 𝑟(𝑦𝑈𝑗 ,𝑣 = 1|𝑈𝑗 ).
the data collection phase to train the recommendation model, the core Static Fairness vs. Dynamic Fairness. Static fairness is defined as
of which is to deduce user preferences from past interaction data to
providing a short-term fair recommender system regardless of changes
predict the possibility of users choosing unvisited targets. Models with
in the recommendation environment. Zhang et al. [68] propose that
design flaws (e.g., ranking bias) may further magnify biases in the input
the concept of recommendation fairness proposed by most methods is
data.
static fairness, i.e., the protected group is fixed during the recommenda-
Ranking Bias. Some loss functions can further exacerbate unfairness
tion process. As an example, traditional matrix-based recommendation
during the training of recommender systems. These loss functions affect
strategies are aimed at maximization of users’ immediate gratifica-
the predicted score of the item, which in turn affects the recom-
tion, assuming that their preferences remain static. However, they
mendation ranking list of recommender systems. For example, Wan
potentially ignore users’ long-term interests.
et al. [63] demonstrate that point losses (e.g., MSE loss) and pair-
Dynamic fairness [27] is defined as considering dynamic factors
wise losses (e.g., BPR loss) are sensitive to popular items. These loss
in the environment to maintain fairness. Dynamic fairness is related
functions give popular items higher scores than unpopular items, which
to time bias. Some works argue that recommending an item that was
amplifies exposure bias during the training of recommender systems.
recommended a long time ago should have the same probability as
Zhu et al. [64] also prove that the BPR loss lacks the fairness constraint
recommending an item that was recommended recently. Assuming 𝑣𝑡
of equal opportunity ranking for reducing bias.
represents an item appears in time 𝑡 and 𝑣𝑡+1 represents an item appears
Feedback Bias. A feedback loop exists in every recommender sys-
in time 𝑡 + 1, the dynamic fairness can be defined as: 𝑃 𝑟(𝑦𝑢,𝑣𝑡 = 1|𝑣𝑡 ) =
tem, which provides recommendation results of a model to users for
𝑃 𝑟𝑡 (𝑦𝑢,𝑣𝑡+1 = 1|𝑣𝑡+1 ). In real-world recommender systems, it can also be
selection. In the feedback loop, users and a recommender system 𝑃 𝑟(𝑦𝑢,𝑣𝑡 =1|𝑣𝑡 )
interact and co-evolve. Users’ preferences and behaviors are updated expressed as 𝑃 𝑟(𝑦𝑢,𝑣 =1|𝑣𝑡+1 )
< 𝜉, where 𝜉 is a slack factor used to adjust
𝑡+1
through the recommender system, and the recommender system uses the dynamic fairness granularity for recommender systems.

5
D. Jin et al. Information Fusion 100 (2023) 101906

Table 1
Summary of methods. We classify existing methods based on the phase of implementation, the technology utilized, the bias issues encountered, the suggested
application scenarios, and the official method names.
Stage Technologies Bias Background Methods
Popularity Bias General RSs IFNAa [69]
Data re-labeling
Pre-processing User Bias General RSs DPTCa [70]
Data re-sampling User Bias General RSs MPMLa [71], HPO [72], RNS [73], HFDa [74]
Data modification User Bias General RSs HURRa [75], CFAIa [76], RSNMF [77], EPTBa [78]
Cold-start Bias Cold-start RSs CLOVER [79]

Regularization Exposure Bias Social RSs MinDiff [80], FRRPCa [81], SERec [82], IDLRa
[83]
Popularity Bias General RSs FARLa [84], CPBLRa [85]
User Bias General RSs CFIFDa [86], FFFUa [87], SLIM [88], FaiRecSys
[89], MMIURa [64], F2VAE [90], FairRec [91],
IURPFa [92], FRRPa [81], FATBRa [93]
Cold-start Bias Cold-start RSs CLOVER [79]
In-processing
Exposure Bias Social RSs DHRSa [94], ABPULa [95]
Causal inference Popularity Bias General RSs IANPa [96], PDA [97], MACRa [98], DANCER [99]
Content RSs FairTED [100]
User Bias
General RSs CSBRSa [101], SHT [102], RTDLEa [103], TPFBa
[104], AdaRequest [105], InvPref [106],
DeSCoVeR [107]
Session-based RSs AIPBa [20]
Adversarial learning User Bias General RSs FairRec [91], FRFC [108], CGWGa [109]
Exposure Bias Social RSs FCPO [27], MORL [110]
Reinforcement learning
Popularity Bias Conversational RSs Popcorn [111]
Exposure Bias Social RSs RDC [112]
Ranking
Popularity Bias General RSs CPR [63]
User Bias General RSs MMIURa [64]
Exposure Bias Session-based RSs CLRec [113], SAR-Net [114]
Others
User Bias Conversational RSs CUABa [56]
User Bias Social RSs FairSR [115]
Cold-start RSs GEN [62]
Non-parametric re-ranking Exposure Bias
Post-processing
General RSs CPFair [116], FairMatch [117], TFROM [118]
Parametric re-ranking Exposure Bias General RSs HyperFair [119]
a Indicates the method has no specific name. We named it with the abbreviation of its article name.

Single-sided Fairness vs. Multi-sided Fairness. There are multiple • Causal Fairness. Most recommender systems are based on statis-
stakeholders in a recommender system. Single-sided fairness refers to tics; however, this ignores causality in the original data. Unlike
maintaining the interests (i.e., fairness) of a single role. For example, individual fairness and equal opportunity, causal fairness not
when recommending items to users, it only pays attention to whether only considers observational data but also incorporates additional
the user has received a fair recommendation. Multi-sided fairness refers causal relationships within them [97]. Counterfactual fairness
to maintaining the interests of multiple roles [116]. For example, con- is a typical causal fairness. It demands that the predictions of
sumers expect the recommender system to fairly recommend products recommender systems remain consistent in both the factual and
they are interested in, while suppliers wish their products to be fairly the counterfactual world [23,124].
exposed to consumers. Multi-sided fairness aims to maintain the fair- • Long-term Fairness. Long-term fairness is affected on time [110]
and considers the long-term impact of fairness interventions [125].
ness of both consumers and suppliers. A typical multi-sided fairness is
It is similar to dynamic fairness in that it considers the long-term
to increase diversity [120] in recommendation results, ensuring that the
impact on user preferences [126].
average exposure of items across multiple aspects is balanced [121]. In
• Envy-freeness Fairness. Do et al. [127] propose a criterion of envy-
addition, it is plausible to enforce proportional exposure of both groups
∑|𝑉𝑖 | ∑|𝑉𝑗 | freeness fairness, stating that each user should prefer their own
𝑃 𝑟(𝑦𝑢,𝑉 =1|𝑉𝑖 ) 𝑃 𝑟(𝑦𝑢,𝑉 =1|𝑉𝑗 )
relative to their average utility, i.e., 𝑖=1 |𝑉 | 𝑖 = 𝑗=1 |𝑉 | 𝑗 , recommendations over those of other users.
𝑖 𝑗
where 𝑉𝑖 and 𝑉𝑗 are two different item groups. Other fairness concepts, like Rawlsian Maximin Fairness or Maximin-
Others. Other common perspectives of fairness in recommender Shared Fairness, can be found in recent literature on fairness [14].
systems are:
3. Methods for fairness-aware recommender systems
• Ranking Fairness. The fairness of sorting is usually reflected in
statistical equality or equality of opportunity. A simple inter- In this section, we summarize methods for improving the fairness
pretation of statistical parity in ranking is the assurance that of recommender systems and present them in three stages of imple-
the proportion of protected individuals appearing within a rank- mentation, including pre-processing, in-processing, and post-processing
ing prefix exceeds a predetermined threshold [122]. In general, (as shown in Fig. 3). In addition, we summarize the fairness-aware
ranking fairness asks that similar items or groups of items re- recommendation methods in Table 1, specifically including the biases
ceive similar visibility, they appear at similar positions in the the methods can address, the types of fairness-aware recommendation
ranking [123]. methods, and other important information.

6
D. Jin et al. Information Fusion 100 (2023) 101906

Fig. 4. Regularization-based fairness-aware methods.

3.1. Pre-processing methods for fairness recommender systems model user preferences as a user–item rating matrix. Since it is not
possible for the user to interact with all the items. The rating matrix is
In the pre-processing stage, fairness-enhancing methods are used to thus high-dimensional and sparse (HiDS), with many missing data rep-
minimize biases in the training datasets. It is possible for these biases resenting the user’s unobserved preferences. Some works [77,78] apply
to be amplified throughout the lifecycle of a recommender system, a latent factor-based model, which provides a good job of handling the
resulting in unfair recommendations for users. We divide these fairness- high-dimensional and sparse (HiDS) matrices, to process missing data.
aware pre-processing methods into three categories according to data
debiasing methods.
3.2. In-processing methods for fair recommender systems
3.1.1. Data re-labeling
Re-labeling changes the labels of the training dataset to remove Currently, most recommender systems (e.g., collaborative filtering
biases in the input data. Some recommender systems predict user methods [96]) are devised to extract user preferences by learning the
preferences using binary implicit feedback labels, like ‘‘click’’ or ‘‘not correlation in the training data [102]. However, people potentially
click’’. However, there may be noise in the feedback labels, i.e., clicks suffer from unfairness services when using these correlation-oriented
do not necessarily represent positive feedback, and missing clicks do recommender systems, such as Simpson’s paradox [128], popularity
not necessarily represent negative feedback. These ‘‘noisy’’ labels can bias [97], user-oriented bias [105], cold-start bias [79], to name only
lead to a drop in model performance. Wang et al. [69] design a self- a few. In addition to removing bias via pre-processing methods, many
supervised re-labeling framework for noise in implicit feedback. The works design fairness-aware methods to alleviate or even eliminate un-
framework dynamically generates pseudo labels for user preferences to fairness during model training of recommender systems. In-processing
mitigate noise in both observed and unobserved feedback data. fairness-aware methods aim to learn bias-free models. In this section,
we classify these works into five categories including regularization-
3.1.2. Data re-sampling based methods, casual-inference-based methods, adversarial-learning
Unbalanced data distribution may lead to a decrease in the training based, reinforcement-learning-based methods, ranking methods, and
effect of recommender systems. Montanari et al. [72] design a data others.
sampling algorithm to ensure that the sampling is uniform and not
affected by the distribution. They randomly select a certain percentage
of users and delete their interaction information. This type of re- 3.2.1. Regularization and consternation for fairness
sampling method reduces the size of the dataset for approximately To alleviate unfairness in recommender systems, regularization pe-
the same proportion of users and maintains the dynamic nature of nalizes the predicted recommendation score in accordance with the
user profiles. Celis et al. [74] propose a subsampling method that fairness evaluation [34]. Regularization terms are widely used in var-
proportionally subsamples the dataset based on different sensitive at- ious recommendation scenarios to reduce bias, such as user-oriented
tributes. Ding et al. [73] design a negative sampler that creates data bias [79], group bias [81], exposure bias [82], popularity bias [84],
resembling generates data similar to the exposure data through feature- etc.
matching techniques instead of selecting directly from exposure data. In general, a regularization term is usually regarded as an additional
The sampler forces the distribution of positive and negative data to be loss focusing on promoting fairness. The term is added to the main loss,
balanced by adding negative samples. which is mainly responsible for improving the recommendation perfor-
mance. Specifically, the total loss used for recommendation training is
3.1.3. Data modification the sum of the performance loss and the regularization term:
The main idea of data modification is to augment or modify the
biased data to reduce the bias. For example, some recommender sys-  = 𝑝𝑒𝑟 (𝐮, 𝐯) + 𝑟𝑒𝑔 (𝜽), (2)
tems may utilize textual data (e.g. job recommender systems, news
where 𝐿 is the total loss for training, 𝐿𝑝𝑒𝑟 (⋅) is the loss for optimizing
recommender systems), which may be missing or have errors. This
the recommendation performance, 𝐿𝑟𝑒𝑔 (⋅) is a customized loss function,
incomplete or noisy data can lead to data bias. To solve the noisy
data problems, Wang et al. [76] implemented natural language pro- and 𝜃 refers to all possible parameters related to fairness evaluation.
cessing (NLP) pre-processing techniques to modify the training data. Fig. 4 shows a general framework for regularization-based fairness-
Specifically, in a crowdtesting recommendation (e.g., recommending aware methods. According to the composition of the regularization
software testing tasks to professionals), they first perform standard term, we can divide the regularization into three categories, including
word segmentation for each document. Then, they remove stop words norm-based regularization terms, matrix-based regularization terms,
and apply synonym substitution to reduce noise. In addition, they and pair-wise regularization terms.
construct a descriptive term list and perform term filtering for each Norm-based Regularization. The norm calculates the distance
document. Similarly, Sachdeva et al. [75] use a similar NLP approach between raw features and generated embeddings. It is used to evaluate
as Wang et al. In addition to handling noisy data, some works focus the deviation of the model of learnable recommender systems. Burke
on missing data. Collaborative filtering-based recommenders typically et al. [88] propose the possibility of using 𝑙1 norm, and 𝑙2 norm as

7
D. Jin et al. Information Fusion 100 (2023) 101906

Fig. 5. A causal graph is a directed acyclic graph, wherein each node stands for a random variable, and the directed edges indicate causal relationships. In a recommender system,
the recommendation process can be briefly described as calculating the matching score 𝑟 between user 𝑢 and item 𝑣, to predict the recommendation score 𝑌 . As shown in (a),
where 𝑢 is the user’s representation, 𝑣 is the item’s representation, ℎ represents the user’s history, and 𝑦 is the prediction score. 𝑢 → 𝑟 denotes that 𝑢 is an inducement of 𝑟, and
𝑅 has been effected by a direct causal from 𝑢. Similarly, there is no arrow between 𝑢 and 𝑦, indicating that 𝑢 has no direct causal relationship with 𝑦. The popularity bias, a
prevalent form of unfairness in existing recommender systems, results from users repeatedly clicking recommended items and then recommender systems always advocating for
these items. As shown in (b), the history ℎ influences 𝑢 and 𝑣, which determine 𝑟. (c) shows a method to decrease the effect of ℎ on 𝑢 and 𝑣.

regularization terms. Among multi-sided unfairness issues, user neigh- where 𝐮𝑏 and 𝐮𝑑 are the bias-aware and bias-free embeddings, re-
borhoods can constrain deviance in their opinion. The regularization spectively. For group fairness, Boratto et al. [92] customized a special
term can be described as: regularization term to reduce the sensitive attribute in group fairness.
∑𝑛
𝜆 ∑ 2
𝑛
𝜆2 𝑓 (𝑢, 𝑣) ⋅ (𝑣)
𝑟𝑒𝑔 = 𝜆1 ‖𝑊 ‖1 + ‖𝑊 ‖2 + 3 (𝑏 ) , (3) 𝑟𝑒𝑔 = ( 𝑖 ∑𝑛 − 𝐶)2 , (7)
2 2 𝑖 𝑖 𝑖 𝑓 (𝑢, 𝑣)

where 𝑊 is a user–user weight matrix, ‖ ⋅ ‖1 is 𝑙1 norm, ‖ ⋅ ‖2 is 𝑙2 norm, where (𝑣) represents the percentage of users who have interacted with
and 𝑏𝑖 is a neighborhood balance regularization for reducing the prob- item 𝑣, 𝐶 represents the proportion of interactions between groups
ability of user neighborhoods forming. 𝑏𝑖 is the squared difference be- with sensitive attributes and a certain type of item in all interactions.
tween the weights of the protected users versus the unprotected users. This regularized optimization implies that the model is penalized if the
Protected users are usually including sensitive attributes, and unpro- difference between correlation and contribution of the population is
tected users do not have sensitive attributes. The works [87,93] employ significant. Zhu et al. [64] use Kullback–Leibler Divergence to normal-
similar techniques to alleviate group unfairness. Hu et al. [86] use the ize user prediction scores to a normal distribution, reducing the ranking
unfairness in model training.
𝑙2 norm of user embeddings and item embeddings as a regularization
term.
3.2.2. Causal inference for fairness
Matrix-based Regularization. Some regularization terms are in
The causal inference in artificial intelligence explores the causal
the form of matrices. Abdollahpouri et al. [85] introduce a matrix-
relationships between variables, i.e., how one variable determines
based regularization LapDQ to reduce the popularity bias. The LapDQ
another variable. In this survey, causal inference for fairness represent
regularizer is defined as :
methods (e.g. inverse propensity score [95,129]) that trace to the
𝑟𝑒𝑔 = 𝑡𝑟(𝑉 𝑇 𝐿𝐷 𝑉 ), (4) source of bias and then mitigate unfairness through causal infer-
ence [98,104,106]. As shown in Fig. 5, causal graphs, which visually
where 𝑉 is the item embedding matrix, 𝑡𝑟(⋅) is the trace function, and represent causal relationships between variables in a recommender
𝐿𝐷 is the Laplacian of the dissimilarity matrix 𝐷. Wasilewski et al. [83] system, are used to analyze the causes of unfairness. In this section,
and Edizel et al. [89] also use this type of regularizer for ranking based on the usage of causal graphs, we categorize current causal in-
unfairness and user bias. ference methods for fairness into inverse propensity scoring, backdoor
Correlation-based Regularization. This type of method mainly adjustment, and counterfactual inference.
exploits correlation to reduce bias. Beutel et al. [81] propose a Inverse Propensity Score. By analyzing the causes of bias in the
correlation-based pair-wise regularization term to balance the clicked causal graph, the inverse propensity score (IPS) reweights samples in
and unclicked item, which is defined as: order to reduce the influences of biased samples, without changing
the causal relationship between variables [94]. In Fig. 5(b), for in-
𝑟𝑒𝑔 = |𝐶𝑜𝑟𝑟(𝐴, 𝐵)|, (5) stance, the confounding variable ℎ affects both 𝑢 and 𝑖, the probability
distribution 𝑃̂𝑟(ℎ, 𝑢, 𝑣, 𝑟) in this figure can be expressed as
where 𝐶𝑜𝑟𝑟(⋅) calculates the absolute correlation of two random vari-
ables, 𝐴 and 𝐵 are two random variables. The specific meaning of 𝑃̂𝑟(ℎ, 𝑢, 𝑣, 𝑟) = 𝑃 𝑟(𝑟|𝑢, 𝑣, ℎ)𝑃 𝑟(𝑢|ℎ)𝑃 𝑟(𝑣|ℎ), (8)
𝐴 in this work is the residuals between clicked and unclicked items.
Variable 𝐵 means the correlation between group users of the clicked while in the absence of the confounding variable (i.e., Fig. 5(c)), the
distribution is:
items. The model is penalized if it predicts that one group clicked on
an item more than the other group did. Prost et al. [80] propose a 𝑃 𝑟(ℎ, 𝑢, 𝑣, 𝑟) = 𝑃 𝑟(𝑟|𝑢, 𝑣, ℎ)𝑃 𝑟(𝑢)𝑃 𝑟(𝑖)𝑃 𝑟(ℎ). (9)
MinDiff formulation based on the above method. MinDiff minimizes
the correlation of predicted probability distribution and the distribution To remove the bias brought by the confounding variable ℎ, IPS methods
between the clicked items and unclicked items. expect 𝑃̂𝑟(ℎ, 𝑢, 𝑣, 𝑟) = 𝑃 𝑟(ℎ, 𝑢, 𝑣, 𝑟), which can be achieved through
multiplying 𝑃̂𝑟(ℎ, 𝑢, 𝑣, 𝑟) by a propensity weight, i.e.,
Others. To further eliminate data bias in the model training, Wu
et al. [91] propose an orthogonality regularization to orthogonalize 𝑃 𝑟(ℎ)
𝑃 𝑟(ℎ, 𝑢, 𝑣, 𝑟) = 𝑃̂𝑟(ℎ, 𝑢, 𝑣, 𝑟). (10)
the unbiased user embeddings to the biased user embeddings. It thus 𝑃 𝑟(𝑢|ℎ)𝑃 𝑟(𝑣|ℎ)
distinguishes between embeddings that are unbiased and those that are Recently, some methods using IPS have been studied to conduct
biased. For each user 𝑢, the orthogonality regularization can be defined debiasing. A propensity framework proposed by [95] makes propensity
as: estimations for improving exposure fairness in implicit feedback sce-
𝐮𝑏 ⋅ 𝐮𝑑 narios. Xiao et al. [129] fuse a deep variational information bottleneck
𝑟𝑒𝑔 (𝑢𝑏 , 𝑢𝑑 ) = | |, (6) approach with a propensity score to develop an unbiased learning
‖𝐮𝑏 ‖⋅ ∥ 𝐮𝑑 ∥

8
D. Jin et al. Information Fusion 100 (2023) 101906

Fig. 6. Adversarial-learning-based fairness-aware methods. The adversarial-learning-based methods first use a generator and a sensitive attribute filter to remove sensitive
information, resulting in embeddings without sensitive information, which are used for recommendation predictions. The corresponding sensitive attributes are predicted by
the discriminator from the filtered embeddings. The generator and the discriminator engage in a max–min game.

algorithm. Since user preferences might drastically change over time, is transformed into the causal graph of the counterfactual world. The
Huang et al. [99] exploit a dynamic inverse propensity score for debi- causal relation 𝐼 → 𝑌 is removed by the following formula to alleviate
asing dynamic popularity biases. To address user behavior bias, Wang the item popularity bias problem:
et al. [101] use unbiased data to introduce propensity scores into biased
𝑦̂𝑟 ∗ 𝜎(𝑦̂𝑖 ) ∗ 𝜎(𝑦̂𝑢 ) − 𝑐 ∗ 𝜎(𝑦̂𝑖 ) ∗ 𝜎(𝑦̂𝑢 ), (11)
training of recommender systems. Another recent work [130] analyzes
the data bias mechanism in the sequential recommendation and re- here the hyperparameter 𝑐 controls the influence of user and item
weights the training parameters to reduce bias using inverse propensity properties on the prediction result. The inference can be interpreted
scores. logically as a ranking adjustment based on 𝑦̂𝑢𝑖 .
Backdoor Adjustment. Unlike the reweighting operation in IPS FairTED [100] creates counterfactual samples of sensitive attributes
methods, backdoor adjustment achieves fairness by blocking off relation- to make sure that the speaker’s sensitive attribute (i.e., gender) cannot
ships that lead to biases. When removing confounding factors, backdoor influence the TED talk quality prediction. Specifically, when generating
adjustment methods require the causal relationship between variables counterfactual samples, the score of presentations with female speakers
satisfies the backdoor criterion. Here, the variable set 𝑍 satisfies the are assigned as the score of presentations with the same contents
backdoor criterion on a causal relationship 𝑢 → 𝑣 in the causal graph, and male speakers. These counterfactual samples are added into the
indicating 𝑍 satisfies (1) there is no descendant node of 𝑢 in 𝑍, and training dataset to develop recommender systems with counterfactual
(2) Every path between 𝑢 and 𝑣 that leads to 𝑢 is blocked by 𝑍. For fairness on gender. To mitigate the popularity bias and improve ex-
example, as shown in Fig. 5(c), the history 𝐻 affects the representation plainable fairness, Ge et al. [132] propose a framework called CEF,
of 𝑢 and 𝑣. To block the influence of ℎ on 𝑢 and 𝑣, backdoor adjustment which uses counterfactual inference (i.e., introducing small changes
methods employ a 𝑑𝑜 operation in cutting off the edge of ℎ → 𝑢 and in the features) to find the root cause of the model’s bias. In CEF,
the edge of ℎ → 𝑣, which can be formulated as 𝑃 𝑟(𝑦|𝑑𝑜(𝑢), 𝑑𝑜(𝑣)) = the scores of each feature, which are calculated from counterfactual
𝑃 𝑟(𝑦|𝑢, 𝑣). recommendation results, are regarded as fairness explanations. More-
Although the backdoor adjustment is effective in removing unfair- over, Li et al. [104] counterfactually infer that user-sensitive features
ness, it faces efficiency challenges resulting from an unlimited sample should be orthogonal to user embeddings, and make fair personalized
space of confounding factors. To this end, Wang et al. [131] intro- recommendations by removing user-sensitive features. In the music
duce DecRS, a model that uses data approximation and KL diver- streaming media recommender system, the user may have a situation
gence to adjust the backdoor criterion. To lessen document-level label where the music is playing incorrectly. Zhang et al. [56] suggest a
bias in text-contained recommender systems, DeSCoVeR [107] uses counterfactual learning strategy to correct user feedback that has been
causal backdoor adjustment and sentence-level keyword bias elimina- incorrectly categorized.
tion techniques in a semantic context. An inference model involving
popularity-bias deconfounding and adjusting (PDA) is proposed by 3.2.3. Adversarial learning for fairness
Zhang et al. [97] as a new inference approach. It employs backdoor Adversarial learning is a method commonly used in recommender
adjustment during model training to eliminate confusion caused by systems to remove sensitive attributes. An adversarial-learning-based
popularity bias. fairness-aware framework generally consists of a generator that pro-
Counterfactual Inference. Counterfactual inference methods for duces node embeddings and a discriminator that predicts sensitive
fairness construct a counterfactual causal graph [98] based on the real features from these generator outputs. By playing a min–max game
causal graph through some fairness-concerning actions (e.g., changing with the discriminator and generator, adversarial-learning methods
the values of some sensitive attributes like gender, age, race [20,21]). are able to learn fair representations. Passing a negative gradient by
Recommender systems are fair and unbiased (i.e., counterfactual fair- predicting the sensitive attribute enables the model to fool the dis-
ness in Section 2) if the recommendation results in the real world and criminator, so that the information content of the sensitive attribute
the counterfactual world are the same. The intuition of counterfactual is continuously reduced. When the discriminator cannot predict the
inference methods can be understood as ‘‘if a fairness-concerning action sensitive feature value, the output of the generator is considered to
(e.g., modifying the value of gender) cannot change recommendation be decoupled from the sensitive feature. The general form of loss for
results, then results from the recommender systems are not affected adversarial-learning-based fairness-aware methods can be formulated
by the action-object (e.g., gender)’’. A representative counterfactual as:
inference method for fairness is the framework called MACR [98],
 = 𝑝𝑒𝑟 (𝑧, 𝑦) + 𝑎𝑑𝑣 (𝑦, 𝑠), (12)
which is proposed to mitigate popularity bias. MACR assumes that the
recommendation interaction matrix 𝐼𝑢𝑣 is affected by user 𝑢, item 𝑣, and where 𝑧 is a set of the generated representations of the generator, 𝑦 is
the user–item matching ranking score 𝑦̂𝑟 , i.e., 𝑦𝑢𝑣 = 𝑦̂𝑟 ∗ 𝜎(𝑦̂𝑣 ) ∗ 𝜎(𝑦̂𝑢 ), a set of predictions, and 𝑠 is a set of the predictions of the discrimi-
𝑦̂𝑣 indicates the influence from item popularity, and 𝑦̂𝑢 represents the nator. Fig. 6 shows a general framework for adversarial-learning-based
extent of the user 𝑢 interact with items. The higher value of the 𝑦̂𝑢 , fairness-aware methods. Recent work [79] is a study addressing cold-
the more likely the user is affected by the popularity of the item. start bias. To swiftly adjust to cold-start new users, it suggests a
Through counterfactual inference, the causal graph of the real world comprehensive fair meta-learning framework (CLOVER) to gather a

9
D. Jin et al. Information Fusion 100 (2023) 101906

Fig. 7. Reinforcement-learning-based fairness-aware methods. A recommender system retains the learned experience in the process of interacting with the environment. When
interacting in the next round, the behavior with the largest positive feedback and the smallest fairness cost will be selected.

general understanding of user preferences. CLOVER establishes that system, modeling the recommendation process as a constrained Markov
recommenders with ratings based on these representations will also ful- Decision Process (CMDP). CMDP proposes two dynamic fairness con-
fill counterfactual fairness if individual fair adversarial games converge straints for reinforcement learning. The first is the population equality
to optimal solutions. Through adversarial learning, CLOVER improves constraint, which requires equal average exposure for each group of
the fairness of the cold-start problem in recommender systems. Based items. This constraint is enforced at all reinforcement learning iter-
on sensitive attributes, Wu et al. [91] craft a bias-aware user em- ations. The second is the exact-k fairness constraint, which requires
bedding, which is specially used to capture the bias. Furthermore, it that the length of protected candidates in each recommendation list
also learns a bias-free user embedding, which is only used to encode is statistically below a given threshold. In interactive recommender
attribute-independent information that the user interests. The fairness systems (IRS), Liu et al. [136] dynamically maintain the long-term
of the model is guaranteed by orthogonalizing the two types of embed- trade-off between accuracy and fairness by offering a method called
dings. Liu et al. [108] propose an adversarial graph neural network FairRec. Using reinforcement learning, recommendations are generated
(GNN) to prevent users from being affected by sensitive features of by combining user preferences with system fairness in FairRec. In
neighboring users. In particular, they propose two fairness constraints addition, FairRec introduces a concept called weighted proportional
to address the failure and inefficiency of adversarial classifiers in the fairness to ensure the fairness of item exposure.
training data. Rus et al. [109] use adversarial learning to mitigate gen- Moreover, some approaches have proposed employing multi-
der bias in word embeddings obtained from recruitment-related text, objective reinforcement learning to improve fairness. For instance, Ge
which aims to provide unbiased job recommendations to job applicants. et al. [110] investigate the Pareto optimal/effective fairness-utility
To alleviate multi-sided fairness. Liu et al. [133] design explicit and trade-off problem in the recommendation process. Using multi-objective
implicit adversarial fairness discriminators. Explicit discriminators aim reinforcement learning, they suggest creating a fairness-aware recom-
to address biases from a local perspective, while implicit discriminators mendation framework (MoFIR), which introduces conditional networks
focus on addressing biases from a global perspective. The fairness and modifies the network according to user preferences. Fu et al. [111]
generator and discriminator are trained adversarially together to miti- propose a multi-objective MDP-based framework, namely Popcorn, for
gate bias. In their paper, Wu et al. [48] propose a fairness awareness eliminating popularity bias in conversational recommender systems.
framework relying on prompts-based bias eliminators in combination Popcorn effectively balances recommendation performance and item
with adversarial training. Li et al. [134] devise a generative adversarial popularity through a real-time semantic understanding of user history
network (GAN), named FairGAN, designed to generate negative signals to avoid long-tail effects.
of users to ensure data fairness. These signals enable FairGAN to
complete the best item exposure ranking. 3.2.5. Ranking optimization for fairness
Ranking is an important part of the recommendation algorithm. The
3.2.4. Reinforcement learning for fairness recommender system recommends items for a user according to the
Recently, some studies take the dynamic interactions among users, ranking of items. Specifically, the ranking algorithm sorts candidate
items, and recommender systems into consideration, and model this items according to users’ preference, and generates a ranking list. Top-
feedback loop as a Markov Decision Process (MDP) [135]. With this scoring candidates receive the most exposure and are ranked first. The
view, reinforcement learning (RL) methods are used in training rec- top-𝑘 candidates are usually returned. A learnable recommender system
ommendation strategies from users’ historical information to learn usually uses a loss function for ranking. Depending on the flaw in the
their preferences. In this survey, reinforcement learning for fairness in- design, some biases, such as popular bias and exposure bias, can be
dicates methods that introduce fairness-concerning feedback (e.g., dy- amplified in losses with these flaws, which may lead to unfair out-
namic fairness) during the training of RL-based recommender systems. comes. In this survey, ranking optimization for fairness indicates methods
Fig. 7 shows a general framework for reinforcement-learning-based that reduce unfairness in recommender systems by employing unbiased
fairness-aware methods. loss functions. Fig. 8 shows a general framework for ranking-based
Long-term fairness and dynamic fairness are two typical concepts fairness-aware methods.
concerning mitigating time bias (see Section 2), which can be involved There are two types of ranking losses commonly used in a learnable
in the training of RL-based recommender systems. Here we introduce recommender system: point-wise loss and pair-wise loss. A point-wise
some representative reinforcement learning for fairness methods. Ge loss, such as binary cross-entropy (BCE) [137] and mean squared error
et al. [27] majorly focus on the unfairness of item exposure across (MSE) [101], minimizes the difference between the calculated recom-
groups caused by time bias. According to their opinion, fairness con- mendation score and the ground truth to capture user preferences for
straints ought to change over time. For instance, an item may no longer individual items. Pair-wise loss, such as Bayesian Personalized Ranking
be popular at time 𝑡 + 𝑛, but if it is still exposed in accordance with (BPR) [138], maximizes the user preference gap between the observed
the fairness requirements from time 𝑡 earlier, the long-term dynamic items and the unobserved items. The observed items are interacted
changes in fairness are disregarded. To enable the model to dynamically with users and the unobserved items have no interaction with users.
adjust the recommendation strategy and guarantee that the fairness The BPR loss is a pair-wise ranking algorithm, that performs pairwise
requirements are always satisfied with environmental changes, they comparisons of items, and learns the order of relevant pairs from the
propose a fairness-constrained reinforcement learning recommender comparisons to rank. The BPR loss encourages the observed item’s

10
D. Jin et al. Information Fusion 100 (2023) 101906

Fig. 8. Ranking-based fairness-aware methods. Such methods use a carefully designed unbiased loss for ranking.

prediction to be higher than the unobserved item’s prediction. The BPR the recommendation results provided by target models, which are
loss can be summarized as: treated as black-box during the rearrangement (i.e., re-ranking), after
∑ the training of recommender systems. Re-ranking methods include
𝐵𝑃 𝑅 = − ln 𝜎(𝑦𝑢,𝑣1 − 𝑦𝑢,𝑣2 ), (13)
manual-based re-ranking and algorithmic-based re-ranking. Here, we
(𝑢,𝑣1 ,𝑣2 )∈𝐷𝑆
mainly introduce the re-ranking method that helps to boost the rec-
where 𝑦𝑢,𝑣 is the predicted score, 𝐷𝑆 = {(𝑢, 𝑣1 , 𝑣2 ) ∣ 𝐼𝑢,𝑣1 = 1, 𝐼𝑢,𝑣2 = 0}, ommender system fairness, without repeating the manual re-ranking.
and 𝐼𝑢,𝑣 is the interaction between the user 𝑢 and the item 𝑣. 𝐼𝑢,𝑣 = 1 Depending on whether or not the manual intervention is required,
indicates 𝑣 is an observed item, and vice versa. current re-ranking methods can be categorized into manual re-ranking
It can be seen from the above equation that the BPR loss is affected and algorithmic-based re-ranking methods. In this survey, we focus
by the observed item. Popular items tend to have higher observation on the algorithmic-based re-ranking methods and divide them into
than unpopular items, thus, they have a high exposure probability. non-parametric re-ranking and parametric re-ranking, according to if a
The more items are observed by the user, the higher the exposure for parameter learning process is required to conduct re-ranking. Fig. 9
the user, which leads to popularity bias and exposure bias. Flaws in shows a general framework of re-ranking methods for fair recom-
the BPR loss cause the popularity bias to be amplified during training. mender systems.
Therefore, Wan et al. [63] propose a cross pairwise ranking loss 𝐿𝐶𝑃 𝑅
for unbiased training: 3.3.1. Non-parametric re-ranking
∑ 1 In this survey, non-parametric re-ranking methods represent learning-
𝐿𝐶𝑃 𝑅 = − 𝑙𝑛𝜎[ (𝑦𝑢1 ,𝑣1 + 𝑦𝑢2 ,𝑣2 − 𝑦𝑢1 ,𝑣2 − 𝑦𝑢2 ,𝑣1 )], (14)
(𝑢 ,𝑢 ,𝑣 ,𝑣 )∈𝐷
2 free algorithms to conduct re-ranking concerning fairness. Heuristic
1 2 1 2 𝑐
methods used to achieve fairness are generally non-parametric re-
where 𝐷𝑐 = {(𝑢1 , 𝑢2 , 𝑣1 , 𝑣2 ) ∣ 𝐼𝑢1 ,𝑣1 = 1, 𝐼𝑢2 ,𝑣2 = 1, 𝐼𝑢1 ,𝑣2 = 0, 𝐼𝑢2 ,𝑣1 = 0} ranking strategies, which treat the re-ranking process as an integer
denotes the training data, and 𝐼 is the user–item interaction set. The programming problem. Under specified fairness constraints, the op-
CPR loss uses cross pair-wise interactions as training samples. Given timal re-ranking outcomes are then found through heuristic search
two users and their interacted items, the unobserved data is obtained by methods. Specifically, given original top-𝑘 recommendation results for
exchanging items for the user. CPR decomposes exposure probabilities each user from recommender systems, heuristic methods are designed
of items into a user-specific, item-specific propensity and user–item to maximize the total preference score with respect to fairness by
relevance, which are not independent of each other and thus contain rearranging original recommendations, which can be formulated as
bias. These predicted scores expressed by exposure biases can cancel
the exposure biases each other according to Eq. (14), therefore the CPR ∑
𝑛 ∑
𝑁
arg max 𝐼𝑢𝑣 𝑆𝑢𝑣 ,
loss is unbiased. There are also ranking methods that analyze the flaws 𝐼𝑢𝑣
𝑢=1 𝑣=1
of BPR loss [64,81]. However, these methods achieve unbias ranking (15)
𝑠.𝑡.𝐶(𝑔1 , 𝑔2 ) < 𝜉,
by adding regularization terms, so we will not go into details here.
𝐼𝑢𝑣 ∈ [0, 1],
3.2.6. Others where 𝐼𝑢𝑣 is an interaction matrix that reflects whether item 𝑣 should
In addition to the above-mentioned typical methods, there are be recommended to user 𝑢, 𝐶(⋅) represents a fairness constraint func-
also some niche methods to enhance the fairness of recommender tion, 𝑔1 and 𝑔2 are two different groups. The objective function is to
systems. Zhou et al. [113] theoretically demonstrate that contrastive maximize the recommendation score under fairness constraints, which
loss can replace the inverse propensity score to reduce exposure bias. is composed of user–item matching and user preference. By treating the
Shen et al. [114] learn cross-scenario user interests through an atten- optimization as a 0–1 integer programming problem, a series of works
tion network, and propose a fairness factor to gauge how important employ heuristic algorithms to conduct re-ranking. Here we present
each scenario is. Zheng et al. [53] design a context-bias-aware rec- some representative methods.
ommendation model. They use attention networks to infer negative To ensure fairness across various groups, Fu et al. [139] utilized
user preferences and eliminate contextual bias caused by the combined a heuristic re-ranking algorithm that constructs bias-constrained ex-
interaction between multiple items. Li et al. [115] design an end-to-end plainable recommendations on knowledge graphs. Similarly to this, Li
model for time-influenced sequential recommendation by weighting the et al. [15] proposed a similar approach, employing a heuristic-based
embeddings to mitigate the unfair distribution of user attributes over method for re-ranking users, aiming to enhance group fairness consid-
items. ering their attributes. According to Wu et al. [118], there is a model
called TFROM, which is designed to enhance exposure fairness for both
3.3. Post-processing methods for fair recommender systems users and providers. In this model, the recommendation list’s length is
equated to the capacity of a knapsack, and the items symbolize the ob-
From the view of fairness, the recommendation results from target jects within this knapsack. TFROM uses a heuristic search algorithm to
systems are usually not optimal, since these results potentially do solve the knapsack problem. Zhu et al. [62] present a novel re-ranking
not take factors concerning fairness (e.g., interactions between items, framework including a score scaling model, which is used to re-rank
manual intentions, and differences in user preferences) into considera- the biased recommendation results. Naghiaei et al. [116] convert the
tion. For enhancing fairness, post-processing methods aim to rearrange recommendation optimization problem into a 0–1 integer programming

11
D. Jin et al. Information Fusion 100 (2023) 101906

Fig. 9. The pipeline of re-ranking methods for fair recommender systems. These methods use recommendation lists and fairness constraints to construct integer programming
problem, and use heuristic algorithms or learnable models to re-rank the recommendation lists.

and knapsack problem, and then propose employing greedy algorithms 4.2. Evaluation
or solvers (e.g., Gurobi) to perform re-ranking. Moreover, Mansoury
et al. [117] transform the item exposure inequity problem in popularity Evaluating a fairness-aware recommender system can be measured
into a graph maximum flow problem to conduct re-ranking. Recent in terms of its accuracy and fairness. To avoid bias in recommenda-
work [140] design a two-stage approach to transform the problem tion results, this section presents accuracy metrics that will not harm
into a max–min problem, using greedy strategies to ensure maximum fairness first, then metrics for assessing the fairness of recommender
systems.
exposure for producers and sorting to ensure fairness for each user.
Accuracy Metrics for Fairness. There are some evaluation metrics
used for evaluating the quality of a model. The common evaluation
3.3.2. Parametric re-ranking metrics are as follows:
In this survey, all re-ranking methods that require a learning process • Area Under Curve (AUC) [156] is the region contained within the
to improve the fairness of recommender systems during the post- Receiver Operating Characteristic (ROC) curve and the coordinate
processing stage are called parametric re-ranking methods. Existing axis. The maximum possible value of AUC is one. The closer the
study [62] shows that, compared with non-parametric re-ranking meth- AUC is to 1, the higher the authenticity of the detection method;
ods, parametric re-ranking methods own better debiasing ability and and vice versa.
are more competent when enhancing equality opportunity-based fair- • Precision (P) [157] is the proportion of correctly classified positive
ness. Here we introduce several typical parametric re-ranking meth- examples to all positive examples that are classified, and Preci-
ods. Given the recommendation results from the target systems, Zhu sion@k (P@k) represents the Precision of the top-𝑘 items in the
et al. [62] employ a learnable autoencoder module to enhance fairness list.
• Recall (R) [158] refers to how many of the positive samples are
and ensure the rearranged results have the same distribution as the
successfully found by the model. Recall@k indicates the recall of
original recommendations.
the first 𝑘 items in the result list.
• Average Precision (AP) [159] calculates the average precision. The
4. Datasets and evaluation for fair RSs higher the value of AP@K, the more relevant items are present in
the top−𝑘 recommendations and the higher these relevant items
are ranked.
In this subsection, we provide a brief overview of the most com- • Normalize Discounted Cumulative Gain (NDCG@k) [160] provides
monly utilized datasets and assessment metrics in the context of different degrees of relevance, ranking the results according to
fairness-aware recommender systems. relevance and weighting them uniformly.

Metrics to Evaluate Fairness. There are some commonly used

4.1. Datasets metrics for measuring the fairness of recommender systems:

• Gini coefficient is an economic indicator that is used to measure

In this section, we will present several datasets frequently employed the income inequality of a country or region [28]. It is also often
in experiments related to the credibility of recommendation systems. used to measure the fairness of recommender systems, which
measures the value inequality of a frequency distribution [27].
Due to the complexity of fairness study in recommender systems, one
The value of Gini coefficient is between 0 and 1, and smaller
specific dataset can be applied in addressing different bias issues.
values represent greater fairness. Its definition is:
For example, the Amazon Review dataset is used in mitigating user
1 ∑∑
𝑛 𝑛
attribute bias, model weights bias, and popularity bias. It is benefi-
𝐺𝑖𝑛𝑖 𝑐𝑜𝑒𝑓 𝑓 𝑖𝑐𝑖𝑒𝑛𝑡 = |𝑒𝑣𝑖 − 𝑒𝑣𝑗 |, (16)
cial to use a variety of datasets to detect unfairness in recommender 2
2𝑛 𝑒𝑣 𝑖=1 𝑗=1
systems and evaluate fairness-enhancing methods. Here we summarize
where 𝑛 indicates items’ number, 𝑒𝑣𝑖 represents the exposure fre-
commonly used datasets in Table 2, demonstrating their basic infor-
quency of item 𝑣𝑖 , and 𝑒𝑣 ̄ donates the average of items’ exposure
mation, and possible scenarios (e.g., exposure bias, and what kind of frequency.
recommender systems can be used). • KL-divergence computes the divergence between two distribu-
In addition to the datasets mentioned in Table 2, some other tions [161], which can be used to measure various biases. Taking
datasets can also be used to mitigate unfairness. For example, Gowalla popularity bias as an example, KL-divergence can measure the
[154], Ctrip [118] can be used to solve the Pairwise bias. CIKM and distribution difference between user history records and popu-
AliEc can be used to solve the user attribute biases [48]. MinD [155] larity in recommendation results [20]. The recommendations are
and COCO [13] can be used to solve user behavior biases. CiteULike fairer when the value of KL-divergence is lower. The expression
and XING [62] can be used to solve cold-start bias. Clothing, Cell- of KL-divergence is
Phones [139] are used to solve Group fairness. Ciao [98] can be used ∑ 𝐷 (𝑣)
𝐷𝐾𝐿 (𝐷1 (𝑣), 𝐷2 (𝑣)) = 𝐷1 (𝑣)𝑙𝑜𝑔 1 , (17)
to solve Feedback bias. 𝐷2 (𝑣)
𝑣∈𝑉

12
D. Jin et al. Information Fusion 100 (2023) 101906

Table 2
Summary of the selected datasets. We present the dataset, detailing its cardinality—specifically the number
of users, items, and edges, its applicability to fairness issues, and associated reference relations.
Datasets #Users #Items #Edges Fairness Issues References
Exposure Bias [118]
Amazon 3,915 2,549 77,328
Training Bias [141]
User Bias [64]
Exposure Bias [113]
Beauty 22,363 12,101 198,502
User Bias [139], [15], [134]
Exposure Bias [116]
Epinion 49,290 139,738 664,828
Popularity Bias [142]
User Bias [143]
Exposure Bias [116], [21], [140], [26]
Lastfm 1,892 17,632 92,834
Training Bias [141]
User Bias [127]
Exposure Bias [116], [26]
MovieLens 943 1,349 99,287
Popularity Bias [111], [99]
User Bias [104], [144]
Cold-start Bias [62], [79]
Exposure Bias [145], [146], [21]
MovieLens 1M 6,040 3,706 1,000,209 Time Bias [27], [115]
Training Bias [141]
Popularity Bias [143], [147], [148]
User Bias [64], [131], [142], [149]
Popularity Bias [147], [150], [148]
Yahoo!R3 5,400 1,000 183,179
Training Bias [141]
User Bias [151], [112], [152]
Exposure Bias [132]

Yelp 25,677 25,815 731,671 Popularity Bias [98], [111]

Training Bias [141]
User Bias [64], [142], [153]

where 𝐷1 (⋅) refers to a distribution of item 𝑣 and 𝐷2 (⋅) refers to a for describing the relationship between prevalence and relevance dis-
fair distribution of item 𝑣. tributions. The article experimentally identifies scenarios for the use of
• Difference mainly considers that if the distance between the two true-positive metrics or false-positive metrics. Ge et al. [27] propose a
recommendation results is less than the fairness constraint co- fairness indicator known as the popularity rate, which represents the
efficient, then the recommendation system is considered to be proportion of popular items in a recommended list in relation to the
fair [15], i.e., total number of recommended items.

|𝑦𝑣𝑖 − 𝑦𝑣𝑗 | < 𝜉, (18) 5. Industrial applications of fair RSs

where 𝑦𝑣𝑖 is a recommender system prediction of item 𝑣𝑖 , 𝜉
is strictness parameter of fairness requirements. If the predic- This session focuses primarily on the application of fairness-aware
tion score of 𝑣𝑖 and 𝑣𝑗 is smaller than the fairness constraint recommender systems to real-world situations. We introduce the impact
coefficient, we consider that the recommender system is fair. of fairness-aware recommender systems on life from application sce-
narios such as e-commerce and finance. Table 3 highlights real-world
In addition to the above metrics, we can also employ other metrics applications of fair recommender systems.
to evaluate fairness-aware recommender systems. For example, Lesota
et al. [20] propose measures to account for prevalence bias from the 5.1. Fairness in E-commerce recommendations
median, various statistical moments, and measures of similarity that
consider the entire prevalence distribution, including Mean, Median, There are many applications of the fair correction recommenda-
Variance, Skew, Kurtosis, Kendall’s 𝜏 rank-order correlation. If a recom- tion system in the area of e-commerce. To address the uneven data
mender system offers suggestions that suit the preferences of a certain distribution across scenarios and the systematic bias of disadvantaged
group of users, but fails to accurately reflect the preferences of another items that emerge in the re-ranking phase, Shen et al. [114] have
user group, it can be deemed unfair. An analytical metric called mis- introduced a solution to the aforementioned challenges in the form of a
calibration is used by Abdollahpouri et al. [148] to measure the degree Scenario-Aware Ranking Network (SAR-Net). The SAR-Net is designed
to which a recommender system responds to the true preferences of to capture a user’s interests across various scenarios using two tailored
the user. If the Hellinger distance [148] between two distributions attention modules. By assessing the importance of individual samples
exceeds a threshold, then it is called miscalibration. According to and adjusting predictions accordingly, the SAR-Net aims to mitigate
Fu et al. [111], two measures are used to measure popularity bias data bias arising from human intervention. Wu et al. [162] introduce
in conversational recommender systems (CRSs): personalized average that in e-commerce websites such as Amazon and JD.com, a feeds-based
recommendation popularity (PARP) and popularity-rank correlation for recommendation has become a mainstream recommendation mode,
users (PRU). Mena-Maldonado et al. [147] present two novel metrics and users can scroll down to view more products recommended by

13
D. Jin et al. Information Fusion 100 (2023) 101906

Table 3
Industrial applications of fairness-aware recommender systems. We delineate the applications by outlining their
respective domains of existence, the prevalent bias that they grapple with, and the methodological approaches they
adopt to counteract these biases.
Domains Sub-domains Bias Methods Reference
Exposure Bias Ranking [114]
E-commerce
User Bias Re-ranking [162]
Exposure Bias Re-ranking [163], [164]
Education
Popularity Bias Ranking [165]
Job Exposure Bias Counterfactual learning [59]
recommendation
User Bias Adversarial learning [109]
User Bias Regularization [155]
News
Social activities recommendation User Bias Adversarial learning [53]
Popularity Bias Adversarial learning [166]
Popularity Bias Ranking [167]
Streaming media
Time Bias Backdoor adjustment [168]
recommendation
Popularity Bias Re-ranking [123]
User Bias Adversarial learning [169]

feeds. In the rolling feeds recommendation, four grids with pictures are 5.3.2. News recommendations
usually played on the same mobile phone screen, and the similarity News recommendation systems mainly recommend news to users
of the four products will affect the user’s judgment, thus introducing on digital news sites. Qi et al. [155] present ProFairRec, a news recom-
bias. The article considers this phenomenon to be a contextual bias. The mendation framework that prioritizes provider fairness. By integrating
authors propose an unbiased counterfactual learning method to elimi- adversarial learning, the framework ensures that representations of
nate contextual bias. The proposed method is applied to a real-world fair news from providers remain unbiased during the recommenda-
e-commerce website, JD.com. tion process. They suggest the use of orthogonal regularization of
provider-fair and biased representations to decrease the bias associated
with news providers. Zheng et al. [53] argue that the contextual bias
5.2. Fairness in education recommendations
among news items may not be fully captured due to interactions
among multiple items. To address this, they propose a novel context-
Users are hoped to receive an education without bias in a fair bias-aware recommendation model aimed at eliminating context bias
education recommender system. Gómez et al. [163] believe that the and achieving fairness in recommendations. Qi et al. [166] propose
geographic location of teachers has a strong impact on visibility and user encoders with popularity awareness to eliminate popularity bias
exposure. The re-ranking approach overcomes these phenomena by from user behavior and achieve accurate interest modeling. In news
ensuring that each group receives the exposure expected, thereby en- recommendation, several recommender systems utilize multiple heads
suring that different providers are treated fairly. Boratto et al. [165] to capture correlations between news items based on representations
explore how recommender systems in the context of popularity-biased from the news that users view. Yi et al. [171] argue that news click
massively open online courses. A comparison is made of existing al- behavior may also be biased by the way news is presented on on-
gorithms relating to the popularity of courses, catalog coverage, and line platforms. So this paper proposes a bias-aware personalized news
popularity of course categories. Marras et al. [164] provide a formal recommendation approach called DebiasRec. Debiasrec trains a biased
definition of the online education recommendation principle and pro- news recommendation model from biased click behavior and inferring
pose a novel re-ranking method that is conscious of fairness, in an the biased interests of users from the clicked news articles.
effort to strike a balance between personalization and recommendation
opportunities. 5.3.3. Streaming media recommendations
Streaming media recommendation systems include music recom-
5.3. Fairness in social activities mendation, video recommendation, etc. As a means of increasing trans-
parency and fairness to artists in music recommendation systems, Kird-
emir et al. [167] find the presence of video recommendations in
5.3.1. Job recommendations
YouTube’s structural and systematic biases in YouTube. By employing
Job recommendation systems match job seekers with job infor- a graphical probabilistic approach, this study evaluates the structural
mation, and recommend job listings that meet job seekers’ wishes, properties of video recommendations. To eliminate undesired temporal
or lists of talented candidates that meet the requirements of the re- bias, Zhan et al. [168] propose a duration-fault quantization (D2Q)-
cruiter [170]. However, user-sensitive attributes may cause discrimi- based watch-time prediction framework that allows for industrial pro-
nation for users in job recommender systems and reduce users’ trust in duction systems for scaling. The framework has been implemented
them. Several researches are devoted to mitigating discrimination and within the Kuaishou App, a commercial video streaming platform.
improving the fairness of job recommendation systems. Chen et al. [59] This has resulted in a substantial enhancement in predicting real-
find that click-through rates for job advertisements decreased over time video viewing time, thereby significantly improving real-time
time. Thus, they adopt the inverse propensity weighting method and video consumption. A study by Shakespeare et al. [172] examines
customize a new loss function to rank the deviation of ad exposure whether state-of-the-art collaborative filtering algorithms exacerbate or
position. In a recent paper, Rus et al. [109] demonstrate that gender ameliorate artist gender biases. This work designs two methods for
bias can be removed from 12 million job openings and 0.9 million determining why differences are attributed to changes in the distri-
resumes through the use of a generative adversarial network, providing bution of inputs based on gender and user preferences. Melchiorre
fair job recommendations to mitigate the pay gaps between different et al. [173] construct a dataset comprising information about the
genders. music consumption habits and personality traits of Twitter users. This

14
D. Jin et al. Information Fusion 100 (2023) 101906

Table 4
Connections between Fair-aware RSs to Trustworthy RSs. We present various categories of fairness recommendation methodologies associated
with trustworthiness properties, along with a succinct summary of their content.
Trustworthiness Methods for Fair RSs Ref. Sketches
[174] Causal analysis on the relationship between users’ past and future behaviors.
Causal inference
Explainability [132] An explainable weighting method is employed to rank counterfactual
recommendation outcomes effectively.
[175] A counterfactual analysis and explanation are provided to bolster
the effectiveness of explanations and promote fairness in the process.
[139] Presenting a fairness-constrained method that utilizes heuristic re-ranking to
address the issue of unfairness recommendations based on knowledge graphs.
Data modification [19] Constructing antidote datasets to improve the fairness and robustness of
Robustness recommender systems.
Causal inference [176] Utilizing an inverse propensity score helps eliminate polarity bias in group
recommendations, ensuring a more robustness and fairness outcome.
Regularization [177] Improving robustness and control fairness through L2 regularization loss.
Causal inference [79] Training models on risk-free, user-approved privacy data.
Privacy
Others [21] Preventing the exposure of sensitive information within the learned
embeddings.
Discrimination Re-ranking [13] Utilizing re-ranking methods helps minimize discrimination,
promoting fairness and equal representation in the results.
Re-ranking [26] Transforming the issue into a maximum flow problem to improve the
Diversity
diversity.
Regularization [178] Incorporating regularization terms can enhance fairness and diversity,
ensuring a more balanced and inclusive outcome.

work analyzes the recommendation algorithms SLIM and EASE Mult- recommendation models. Therefore, explainability is crucial for
VAE. Their research results reveal notable differences in performance building user trust in recommender systems.
between user groups scoring high and low on certain personality traits. • Robustness. The robustness of recommender systems refers to
The recent work [123] proposes a fairness-conscious re-ranking frame- their ability to continue to operate normally when threatened or
work for quantifying and mitigating algorithmic bias due to data bias. attacked. It requires the recommender system to be safe, reliable,
In an online A/B test of representative rankings of LinkedIn Talent and robust enough to handle errors or inconsistencies in all life cy-
Search or recommendations, the authors propose a strategy aimed at cle stages of the recommender system. Recommendation systems
distributing ranking outcomes according to one or more safeguarded generally rely on user history records to build algorithm models.
attributes, with the goal of achieving fairness principles like equal User history records contain a lot of junk data generated by
opportunity and population parity. This large-scale deployment of a systems and humans, which may lead to a decrease in the fairness
framework that deploys LinkedIn Recruiter to ensure fairness in the and accuracy of the model. Thus, a trustworthy recommendation
recruitment space without impacting business metrics has the poten- algorithm must be robust.
tial to positively impact over 630 million LinkedIn members. Wu • Privacy. Personal data collected by a recommender system
et al. [169] use adversarial learning to reduce bias arising from user- should be safe and able to protect personal privacy. Private
sensitive attributes. Furthermore, they utilize KL divergence to capture data and privacy must be protected throughout the life cycle
less candidate-aware bias. of the recommender system. Private data encompasses both the
information shared by the user and the information derived from
6. Connections with other trustworthy dimensions the user’s interactions with the system. A trustworthy recommen-
dation system should take responsibility for preventing unlawful
Recommendation systems are one of the most crucial parts of the and unfair discrimination against users as a result of the collected
lives of people today. However, some recommendations lack moral data.
basis and restraints, which undermines user confidence and possibly
A recommender system may contain a variety of untrustworthy
transgresses the law. A crisis of trust in the recommendation system can
issues, and biases that cause unfairness may also cause other untrust-
be sparked by a significant volume of biased training data or biased
worthiness. For example, privacy and robustness issues caused by data
recommendation algorithms. Thus, it is uttermost important to have
bias (age and gender). The issue of fairness may arise simultaneously
trustworthy recommender systems. Ideally, a trustworthy recommender
with other trustworthy issues in a recommender system. We are con-
system should be open and transparent, and its methods for obtaining
cerned with the fairness of the recommender system and introduce
results should be explainable. The trustworthiness of a recommender
other trustworthy ethical principles based on the fairness content.
system is mainly evaluated based on four ethical principles, including
We describe the connection between fairness and each trustworthy
explainability, robustness, privacy, and fairness. Here we introduce
property below. Table 4 outlines the connections between fairness and
other ethical principles beyond fairness.
other trustworthiness properties.
• Explainability. Explainability requires that the decisions of the
recommender system should be understandable by people. Specif- 6.1. Connections with explainability
ically, it requires that the decision-making process, and input and
output relationships of recommender systems should be logically Some methods for mitigating fairness issues can also be added to
explained. However, most of the current recommendation models the explainability of models. For example, methods based on causal
operate in the form of ‘‘black boxes’’, sometimes it is not always inference analyze the causes of bias and provide explanations for the
possible to explain why a recommender system produces a par- decision-making process of recommendation models. Xu et al. [174]
ticular output or decision, which can lead to users’ distrust of argue that the explainability of recommender systems involves causal

15
D. Jin et al. Information Fusion 100 (2023) 101906

Table 5
Future directions for fairness-aware recommender systems. We categorize future directions by segmenting them into distinct trajectories,
elucidating the current deficiencies inherent in the fairness-aware RSs, the potential challenges to be encountered, and elucidating the advantages
of resolving these issues.
Directions Current shortcomings Future challenges Challenges-solving benefits
Various fairness concepts Create customized fairness Standardize industry concepts
Concepts
exhibit both distinctions and concepts for diverse while promoting fairness in
interconnected characteristics. recommendation scenarios. diverse scenarios.
There is not a one-size-fits-all Apply suitable fairness Ensuring the most effective
Frameworks
framework for addressing methods to specific fairness solutions are applied.
fairness concerns. issues.
Balancing fairness can Finding a balance between Ensuring optimal results while
Trade-off
sometimes affect accuracy, fairness and recommendation promoting equitable treatment
causing imperfect outcomes. performance. for all users.
Fairness-trustworthiness Explore the interaction rules Foster trustworthy properties in
Trustworthiness
interactions are underexplored between fairness and fairness RSs to increase
in current researches. trustworthiness. trustworthiness.

analysis between the previous and future behaviors of users, which is used extensively for personalization and when protected private data
bound to answer counterfactual questions. An example of the question attributes like gender and ethnicity are misused. A bipartite graph is
can be ‘‘What would happen if a different set of items were purchased’’. typically created organically by users and items in recommender sys-
Counterfactual inference provides a fair framework for recommender tems. Directly abusing some user–item representations will nevertheless
systems, where the constructed counterfactual world explains why the cause the leakage of user-sensitive information, even if user–item in-
model makes the output decisions. Therefore, counterfactual reasoning teractions do not contain any user-sensitive data. This is because user
can simultaneously promote the explainability and fairness of recom- behavior and attributes are found to be correlated in social theory.
mender systems. Ge et al. [132] propose a counterfactual explainable For example, a person’s privacy (such as his gender) can be inferred
fairness framework for group fairness. Specifically, they propose an from his actions. Each user’s embedding has a hidden connection to
explainable weighting method to rank the counterfactual recommen- the behavior of similar users and users who have the same items, in
dation results, which can be seen as an explanation for the final addition to being tied to the user’s behavior.
recommendation. Cornacchia et al. [175] propose a model that fuses Similarly, Wu et al. [21] make an effort to keep user privacy
natural language processing and counterfactual inferencing to provide and sensitive information hidden from the recommender system. They
recommendations for the loans domain. This model provides users with transform the fairness-aware recommendation problem into learning
fairness and transparent advice. The path-based method is a common fair user and item representations, and provide a GNN method (FairGo)
method to make improvements to recommender systems’ explainabil- to avoid any sensitive information from being revealed from the learned
ity. Fu et al. [139] make improvements on user–item path distribution embeddings. To make a fair recommendation and prevent the spread
and fairness of the recommender system by designing a fairness-aware of high-level sensitive information, FairGo designs an ego network,
ranking algorithm. which is user-centric and links the purchased products of the user
and the item. Fairgo designs an aggregation algorithm that prevents
6.2. Connections with robustness high-order information propagation in the ego network and achieves
representation fairness. The ego network embedding and user–item
Robustness can be reflected in the ability to defense attacks on data embedding are mapped into the same space after learning the embed-
and models. Data bias cause unfair, it is also vulnerable to attack, which ding. In this space, filters of sensitive information are used for filtering,
makes the model less robust. Fang et al. [19] propose to construct and finally, fairness training is performed through graph adversarial
antidote data that mimics the rating behavior of users to mitigate
learning. In a cold-start situation, where user–item interaction data
data bias. These data are not considered anomalous data for attacking,
is lacking, recommender systems leverage the data trained from non-
thus it can make an improvement on the model’s robustness. The
cold-start scenarios as the proxy for cold-start user–item information.
same method also be adopted by Rastegarpanah et al. [87]. Dokoupil
However, this approach can leak the privacy of non-cold-started user–
et al. [176] argue that recommender systems should utilize as much
items. Wei et al. [79] suggest training risk-free, user-approved private
unbiased data as possible, whereas real-world training data is biased. In
data, and then making privacy-preserving fair recommendations to
this case, to make the recommender system robust, they use an inverse
cold-start users.
propensity score to remove polarity bias in group recommendations.
As a result, the fairness of group recommendations has been improved.
Zhu et al. [177] design a local model and a global model to improve 6.4. Connections with others
robustness and control fairness through l2 regularization loss. They
improve model robustness and fairness through continuous gradient Discrimination occurs in an untrustworthy recommender system.
optimization. Mansoury et al. [179] propose three different user profile features, and
analyze the possible connection between these features and the differ-
6.3. Connections with privacy ent behaviors of the recommender system for different genders. They
introduce the unfairness of the recommender system caused by gender
Training data in recommender systems may contain some user- discrimination, and find that women get less accurate recommendations
sensitive attributes that may be considered private by users. Even if the than males based on their experiments. This phenomenon indicates that
user’s sensitive data is well protected, privacy leakage may occur dur- the recommendation algorithm is unfair to different genders.
ing the interaction with the recommender system. The leaked private An ethical recommendation system cannot discriminate against vul-
information can be maliciously obtained by other users, which brings nerable groups. In addition to protecting user privacy from the stand-
data bias to the recommendation system and causes unfairness. Some point of ethical and moral norms, we also need to consider the require-
works [51,79] establish a connection between fairness and privacy. ments of relatively underprivileged groups. A number of measures are
Recommender systems may be unfair if users’ private information is suggested by Leonhardt et al. [180] as ways to quantify the influence

16
D. Jin et al. Information Fusion 100 (2023) 101906

Table 6
A comprehensive overview of recommender systems that are both fairness-aware and incorporate trustworthiness features. The horizontal axis
represents the recommendation methods promoting fairness, while the vertical axis corresponds to the trustworthiness features. The blank
areas indicate an absence of related research in those particular domains. These unexplored areas also present opportunities for future research
directions.

Explainability [174], [132] [139]

Robustness [19], [87] [177] [176]
Privacy [21] [79] [51]
Others [178] [13], [26]
Pre-processing Regularization Causal Adversarial Reinforcement Ranking Post-processing
methods inference learning learning methods

of fairness-aware pre-processing techniques on user prejudice. A re- ness-aware recommender systems are broad, including education, soci-
ranking method is developed by Gomez et al. [13] in order to reduce ety, health care, et al. However, each recommendation scenario has dif-
bias caused by the discrimination against teachers’ geographic loca- ferent fairness-aware recommendation methods, such as regularization-
tion. By using black feminist and critical race theory, Schelenz [181] based methods and re-ranking methods. At present, there is no work to
attempts to lessen the unfairness of the user’s political and social analyze which type of fairness method is applicable to a certain type
environment. of fairness issue from the perspective of recommendation scenarios.
In addition to the problem of discrimination, contemporary recom- In addition, due to the diversity of fairness issues, building a unified
mender systems use big data to conduct in-depth and detailed mining recommendation framework to solve all fairness issues can simplify
of users’ historical behaviors, personal characteristics, and other data. the analysis of various fairness issues and also can be quickly applied
They take the learned user’s preference as the main standard and to new scenarios and unknown fairness issues. There is no work yet
provide precise ‘‘Act According to Actual Circumstances’’ recommen- to solve the above problem. We look forward to building a general
dations for the user. While these recommender systems excessively recommendation framework to address different fairness issues in the
future.
collect users’ private data, they limit the possibility of ordinary people
Trade-off between fairness and performance of RSs. As a means
exploring a variety of new fields.
of ensuring the fairness of the recommender system, a system needs
Another form of fairness that some studies suggest is diversity [181].
to reduce the bias in the recommendation output. However, it is chal-
Diversity can alleviate the ethical issues caused by the precise rec-
lenging to combine fairness with accuracy in recommender systems,
ommendation of recommender systems. Mansoury et al. [26] solve
mainly because the goals of fairness and accuracy are inconsistent and
the problem of unfairness by transforming it into a maximum flow
the trade-off between them is substantial. In recommender systems,
problem, which improves the overall diversity and fair distribution of accuracy is determined by the system’s capacity to accurately antici-
recommended items. Sacharidis et al. [178] propose a regularization pate and meet the needs and interests of users. Taking the example
for social recommendations that allows friends to be similar. However, of a multi-stakeholder recommender system (such as suppliers and
within a community, it generally forces members to be more diverse, consumers), users want the recommendation system to recommend
which results in fairer recommendations. products that meet their preferences, and suppliers want to make the
items they provide as fair as possible to recommend to users. If the
7. Future directions fairness of item exposure is maintained, it may lead to a decrease in
the accuracy of recommendations. Therefore, controlling the trade-off
In recent years, attention from the academy and industry communi- between accuracy and fairness becomes critical.
ties has been paid to improving fair recommender systems. However, Mutual promotion with trustworthy properties. At present, build-
thoroughly building fair recommender systems that can be trusted ing a trustworthy recommendation system is the development trend
still faces the following challenges. Table 5 outlines current works’ of artificial intelligence. Credibility includes multiple properties, and
shortcomings, unresolved issues, and potential benefits of resolving fairness is one of them. Most of the current work focuses on the
them. association between fairness and trustworthiness, and does not study
Concepts for fairness. In different recommendation scenarios, the the law of mutual influence between these properties. For example, the
fairness goals that people pursue are also different. For example, in a work of Fu et al. [139] introduces achieving fairness on an explainable
recommender system. This type of work mainly addresses the issue
recommender system with multiple stakeholders, the goal of fairness is
of fairness without improving explainability. In other words, these
to balance the interests of multiple stakeholders. However, the goal of
approaches to fairness do not promote explainability. We hope that
fairness in a time-aware job recommender system is to balance exposure
future works can focus on the intrinsic interaction between fairness and
frequency with old and new job information [59]. The concepts of
other trustworthy properties in recommender systems. In addition to
fairness in different scenarios have both mutually inclusive and differ-
explainability, there is also little work that considers the positive and
ent parts. For example, both long-term fairness and dynamic fairness
negative effects between robustness and fairness.
are fairness affected by time. A statically fair recommender system Most of the current fairness work coexists with only one fairness
may include individual fairness and group fairness. Therefore, it is property, and no work comprehensively considers all fairness proper-
difficult to have a unified and accurate concept to define the fairness of ties. Another challenge in the future is how to integrate fairness and
recommender systems. Obtaining a common concept for different defi- other credible properties in a recommender system, establish internal
nitions of fairness is an important challenge. In addition, there may be correlations between fairness and properties, reduce conflicts between
some different fairness issues in a recommendation scenario, and these properties, and build a credible recommender system. Table 6 shows
fairness issues may be conflicting. A potentially promising approach the current state of research on integrating fairness methods and trust-
is to consider the prioritization of fairness issues in a recommender worthiness properties. It is evident that numerous areas have yet to be
system. Few works consider the importance of fairness issues. This will explored, such as the application of reinforcement learning methodolo-
also be an important challenge in the future. gies to develop explainability and fairness recommender systems. There
General frameworks for fairness. The application scenarios of remains significant research potential in exploring the integration of
fair- trustworthiness and fairness.

17
D. Jin et al. Information Fusion 100 (2023) 101906

8. Conclusion [6] D. Toquica, K. Agbossou, R. Malhamé, N. Henao, S. Kelouwani, M. Fournier,

A recommender system for predictive control of heating systems in economic
demand response programs, IEEE Open J. Ind. Appl. 3 (2022) 79–89.
Recommender systems provide basic artificial intelligence services
[7] M. Bogaert, J. Lootens, D. Van den Poel, M. Ballings, Evaluating multi-label
to facilitate our daily lives. To ensure that fairness-aware recommender classifiers and recommender systems in the financial service sector, European
systems are crucial for people to enjoy trustworthy recommendations, J. Oper. Res. 279 (2019) 620–634.
this survey reviews current efforts on fair recommender systems with [8] T. Hassan, B. Edmison, T. Stelter, D.S. McCrickard, Learning to trust: Under-
the aim of facilitating their implementation and future research. Firstly, standing editorial authority and trust in recommender systems for education,
in: Proceedings of the 29th ACM Conference on User Modeling, Adaptation and
we establish a foundation by introducing fundamental concepts related Personalization, 2021, pp. 24–32.
to recommender systems and fairness. This introduction will later aid [9] A.C. Chuang, N.F. Huang, J.W. Tzeng, C.A. Lee, Y.X. Huang, H.H. Huang,
us in defining fairness in various contexts. Then, we introduce different Moocers: Exercise recommender system in moocs based on reinforcement
types of biases that cause unfairness at each stage of the recommender learning algorithm, in: 2021 8th International Conference on Soft Computing
& Machine Intelligence, ISCMI, IEEE, 2021, pp. 186–190.
system’s lifecycle. Finally, we introduce the manifestations of fairness
[10] A. Agarwal, D.S. Mishra, S.V. Kolekar, Knowledge-based recommendation
in different recommendation scenarios, which paves the way for the system using semantic web rules based on learning styles for moocs, Cogent
introduction of subsequent methods. For fairness-aware recommenda- Eng. 9 (2022) 2022568.
tion methods, we introduce three processing stages: pre-processing, [11] Z. Qin, S.J. Chen, D. Metzler, Y. Noh, J. Qin, X. Wang, Attribute-based
propensity for unbiased learning in recommender systems: Algorithm and case
in-processing, and post-processing. We summarize the categories of
studies, in: Proceedings of the 26th ACM SIGKDD International Conference on
methods and divide them in detail in each processing stage. For the Knowledge Discovery & Data Mining, 2020, pp. 2359–2367.
evaluation of the fairness-aware recommender system, we introduce [12] C. Musto, P. Lops, G. Semeraro, Fairness and popularity bias in recommender
the data sets suitable for different recommendation scenarios and the systems: an empirical evaluation, 2021.
fairness measurement metrics. For the application of fairness-aware rec- [13] E. Gómez, C. Shui Zhang, L. Boratto, M. Salamó, The winner takes it all:
Geographic imbalance and provider (un)fairness in educational recommender
ommender systems in the real world, we describe how fairness-aware
systems, in: Proceedings of the 44th International ACM SIGIR Conference on
recommender systems maintain fairness in areas closely related to Research and Development in Information Retrieval, 2021, pp. 1808–1812.
people’s lives, such as e-commerce, education, and et al. Regarding the [14] Y. Wang, W. Ma, M. Zhang*, Y. Liu, S. Ma, A survey on the fairness of
promotion of human trust in fairness-aware recommender systems, we recommender systems, ACM Trans. Inf. Syst. (2022) http://dx.doi.org/10.1145/
3547333.
start with several trustworthy properties and introduce the correlation
[15] Y. Li, H. Chen, Z. Fu, Y. Ge, Y. Zhang, User-oriented fairness in
between fairness and these trustworthy properties. We conclude this recommendation, in: WWW, ACM / IW3C2, 2021, pp. 624–632.
survey and discuss the future directions of fairness-aware recommender [16] H. Liu, Y. Wang, W. Fan, X. Liu, Y. Li, S. Jain, A.K. Jain, J. Tang, Trustworthy
systems from several novel perspectives, including the development AI: A computational perspective, 2021, CoRR abs/2107.06641.
direction of fair recommender systems, as well as the development of [17] H. Zhang, B. Wu, X. Yuan, S. Pan, H. Tong, J. Pei, Trustworthy graph neural
networks: Aspects, methods and trends, 2022, arXiv preprint arXiv:2205.07424.
fair and trustworthy recommender systems.
[18] W. Fan, X. Zhao, X. Chen, J. Su, J. Gao, L. Wang, Q. Liu, Y. Wang, H. Xu,
L. Chen, Q. Li, A comprehensive survey on trustworthy recommender systems,
CRediT authorship contribution statement 2022, CoRR abs/2209.10117.
[19] M. Fang, J. Liu, M. Momma, Y. Sun, Fairroad: Achieving fairness for recom-
mender systems with optimized antidote data, in: SACMAT, ACM, 2022, pp.
Di Jin: Conceptualization, Project administration, Writing – original
173–184.
draft. Luzhi Wang: Writing – original draft. He Zhang: Writing – [20] O. Lesota, A. Melchiorre, N. Rekabsaz, S. Brandl, D. Kowald, E. Lex, M. Schedl,
original draft. Yizhen Zheng: Writing – original draft. Weiping Ding: Analyzing item popularity bias of music recommender systems: Are different
Writing – review & editing. Feng Xia: Writing – review & editing. genders equally affected? in: Fifteenth ACM Conference on Recommender
Shirui Pan: Supervision, Writing – review & editing. Systems, 2021, pp. 601–606.
[21] L. Wu, L. Chen, P. Shao, R. Hong, X. Wang, M. Wang, Learning fair represen-
tations for recommendation: A graph-based perspective, in: Proceedings of the
Declaration of competing interest Web Conference 2021, 2021, pp. 2198–2208.
[22] E. Holmes, Anti-discrimination rights without equality, Mod. Law Rev. 68
The authors declare the following financial interests/personal rela- (2005) 175–194.
[23] N. Mehrabi, F. Morstatter, N. Saxena, K. Lerman, A. Galstyan, A survey on bias
tionships which may be considered as potential competing interests: Di and fairness in machine learning, ACM Comput. Surv. 54 (2021) 115:1–115:35.
Jin reports financial support was provided by National Science Foun- [24] A.J. Biega, K.P. Gummadi, G. Weikum, Equity of attention: Amortizing
dation. Shirui Pan reports financial support was provided by Australian individual fairness in rankings, in: SIGIR, ACM, 2018, pp. 405–414.
Research Council. [25] D. Pedreschi, S. Ruggieri, F. Turini, Measuring discrimination in socially-
sensitive decision records, in: SDM, SIAM, 2009, pp. 581–592.
[26] M. Mansoury, H. Abdollahpouri, M. Pechenizkiy, B. Mobasher, R. Burke, A
Data availability graph-based approach for mitigating multi-sided exposure bias in recommender
systems, ACM Trans. Inf. Syst. 40 (2022) 32:1–32:31.
No data was used for the research described in the article. [27] Y. Ge, S. Liu, R. Gao, Y. Xian, Y. Li, X. Zhao, C. Pei, F. Sun, J. Ge, W. Ou,
et al., Towards long-term fairness in recommendation, in: Proceedings of the
14th ACM International Conference on Web Search and Data Mining, 2021, pp.
References 445–453.
[28] R. Dorfman, A formula for the gini coefficient, Rev. Econ. Stat. (1979) 146–149.
[1] F. Ricci, L. Rokach, B. Shapira, Recommender systems: Techniques, applications, [29] H. Zhang, X. Yuan, Q.V.H. Nguyen, S. Pan, On the interaction between node
and challenges, in: Recommender Systems Handbook, 2022, pp. 1–35. fairness and edge privacy in graph neural networks, 2023, CoRR abs/2301.
[2] B. Kersbergen, S. Schelter, Learnings from a retail recommendation system on 12951.
billions of interactions at bol.com, in: 2021 IEEE 37th International Conference [30] S. Wu, F. Sun, W. Zhang, X. Xie, B. Cui, Graph neural networks in recommender
on Data Engineering, ICDE, IEEE, 2021, pp. 2447–2452. systems: a survey, in: ACM Computing Surveys, CSUR, 2022, http://dx.doi.org/
[3] Y. Gu, W. Bao, D. Ou, X. Li, B. Cui, B. Ma, H. Huang, Q. Liu, X. Zeng, Self- 10.1145/3535101.
supervised learning on users’ spontaneous behaviors for multi-scenario ranking [31] Y. Deldjoo, T.D. Noia, F.A. Merra, A survey on adversarial recommender
in e-commerce, in: Proceedings of the 30th ACM International Conference on systems: From attack/defense strategies to generative adversarial networks,
Information & Knowledge Management, 2021, pp. 3828–3837. ACM Comput. Surv. 54 (2021) 35:1–35:38.
[4] X. Guo, S. Wang, H. Zhao, S. Diao, J. Chen, Z. Ding, Z. He, J. Lu, Y. Xiao, B. [32] L. Wu, X. He, X. Wang, K. Zhang, M. Wang, A survey on accuracy-oriented
Long, et al., Intelligent online selling point extraction for e-commerce recom- neural recommendation: From collaborative filtering to information-rich
mendation, in: Proceedings of the AAAI Conference on Artificial Intelligence, recommendation, IEEE Trans. Knowl. Data Eng. (2022).
2022, pp. 12360–12368. [33] M. Zehlike, K. Yang, J. Stoyanovich, Fairness in ranking, part ii: Learning-
[5] J. Yang, J. Li, S. Liu, A novel technique applied to the economic investigation to-rank and recommender systems, in: ACM Computing Surveys, CSUR,
of recommender system, Multimedia Tools Appl. 77 (2018) 4237–4252. 2022.

18
D. Jin et al. Information Fusion 100 (2023) 101906

[34] J. Chen, H. Dong, X. Wang, F. Feng, M. Wang, X. He, Bias and debias in [64] Z. Zhu, J. Wang, J. Caverlee, Measuring and mitigating item under-
recommender system: A survey and future directions, 2020, CoRR abs/2010. recommendation bias in personalized ranking systems, in: SIGIR, ACM, 2020,
03240. pp. 449–458.
[35] Y. Ge, S. Liu, Z. Fu, J. Tan, Z. Li, S. Xu, Y. Li, Y. Xian, Y. Zhang, A survey on [65] H. Abdollahpouri, M. Mansoury, R. Burke, B. Mobasher, The unfairness of
trustworthy recommender systems, 2022, CoRR abs/2207.12515. popularity bias in recommendation, in: RMSE@RecSys, CEUR-WS.org, 2019.
[36] S. Wang, X. Zhang, Y. Wang, H. Liu, F. Ricci, Trustworthy recommender [66] M. Naghiaei, H.A. Rahmani, M. Dehghan, The unfairness of popularity bias in
systems, 2022, CoRR abs/2208.06265. book recommendation, in: BIAS, Springer, 2022, pp. 69–81.
[37] M.N. Freire, L.N. de Castro, E-recruitment recommender systems: a systematic [67] M. Marras, L. Boratto, G. Ramos, G. Fenu, Equality of learning opportunity via
review, Knowl. Inf. Syst. 63 (2021) 1–20. individual fairness in personalized recommendations, Int. J. Artif. Intell. Educ.
[38] J. Lu, D. Wu, M. Mao, W. Wang, G. Zhang, Recommender system application 32 (2022) 636–684.
developments: A survey, Decis. Support Syst. 74 (2015) 12–32. [68] D. Zhang, J. Wang, Recommendation fairness: From static to dynamic, 2021,
[39] D. Jin, L. Wang, Y. Zheng, G. Song, F. Jiang, X. Li, W. Lin, S. Pan, Dual intent
CoRR abs/2109.03150.
enhanced graph neural network for session-based new item recommendation,
[69] Z. Wang, Q. Xu, Z. Yang, X. Cao, Q. Huang, Implicit feedbacks are not
in: Proceedings of the ACM Web Conference 2023, WWW 2023, Austin, TX,
always favorable: Iterative relabeled one-class collaborative filtering against
USA, 30 April 2023-4 May 2023, 2023, pp. 684–693, http://dx.doi.org/10.
noisy interactions, in: ACM Multimedia, ACM, 2021, pp. 3070–3078.
1145/3543507.3583526.
[70] F. Kamiran, T. Calders, Data preprocessing techniques for classification without
[40] C. Gao, W. Lei, X. He, M.de. Rijke, T. Chua, Advances and challenges in
discrimination, Knowl. Inf. Syst. 33 (2011) 1–33.
conversational recommender systems: A survey, AI Open 2 (2021) 100–126.
[71] A. Estabrooks, T. Jo, N. Japkowicz, A multiple resampling method for learning
[41] K. Lamertz, The social construction of fairness: Social influence and sense
from imbalanced data sets, Comput. Intell. 20 (2004) 18–36.
making in organizations, J. Organ. Behav. 23 (2002) 19–37.
[42] J. Konow, A positive theory of economic fairness, J. Econ. Behav. Organ. 31 [72] M. Montanari, C. Bernardis, P. Cremonesi, On the impact of data sampling on
(1996) 13–35. hyper-parameter optimisation of recommendation algorithms, in: SAC, ACM,
[43] F.I. Michelman, Property, utility, and fairness: comments on the ethical founda- 2022, pp. 1399–1402.
tions of just compensation law, in: Constitutional Protection of Private Property [73] J. Ding, Y. Quan, X. He, Y. Li, D. Jin, Reinforced negative sampling for
and Freedom of Contract, Routledge, 2013, pp. 117–210. recommendation with exposure data, in: IJCAI, ijcai.org, 2019, pp. 2230–2236.
[44] S. Salloum, D. Rajamanthri, Implementation and evaluation of movie rec- [74] L.E. Celis, A. Deshpande, T. Kathuria, N.K. Vishnoi, How to be fair and diverse?,
ommender systems using collaborative filtering, J. Adv. Inf. Technol. 12 2016, arXiv preprint arXiv:1610.07183.
(2021). [75] N. Sachdeva, J.J. McAuley, How useful are reviews for recommendation?
[45] D. Lian, Y. Wu, Y. Ge, X. Xie, E. Chen, Geography-aware sequential location A critical review and potential improvements, in: SIGIR, ACM, 2020, pp.
recommendation, in: Proceedings of the 26th ACM SIGKDD International 1845–1848.
Conference on Knowledge Discovery & Data Mining, 2020, pp. 2009–2019. [76] J. Wang, Y. Yang, S. Wang, J. Hu, Q. Wang, Context- and fairness-aware in-
[46] A. Ferraro, X. Serra, C. Bauer, Break the loop: Gender imbalance in music process crowdworker recommendation, ACM Trans. Softw. Eng. Methodol. 31
recommenders, in: Proceedings of the 2021 Conference on Human Information (2022) 35:1–35:31.
Interaction and Retrieval, 2021, pp. 249–254. [77] X. Luo, M. Zhou, Y. Xia, Q. Zhu, An efficient non-negative matrix-factorization-
[47] A. Giabelli, L. Malandri, F. Mercorio, M. Mezzanzanica, A. Seveso, Skills2job: based approach to collaborative filtering for recommender systems, IEEE Trans.
A recommender system that encodes job offer embeddings on graph databases, Ind. Inform. 10 (2014) 1273–1284.
Appl. Soft Comput. 101 (2021) 107049. [78] Y. Yuan, X. Luo, M. Shang, Effects of preprocessing and training biases in
[48] Y. Wu, R. Xie, Y. Zhu, F. Zhuang, A. Xiang, X. Zhang, L. Lin, Q. He, latent factor models for recommender systems, Neurocomputing 275 (2018)
Selective fairness in recommendation via prompts, in: Proceedings of the 2019–2030.
45th International ACM SIGIR Conference on Research and Development in [79] T. Wei, J. He, Comprehensive fair meta-learned recommender system, in:
Information Retrieval, 2022, pp. 2657–2662. Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and
[49] A. Datta, M.C. Tschantz, A. Datta, Automated experiments on ad privacy Data Mining, 2022, pp. 1989–1999.
settings: A tale of opacity, choice, and discrimination, 2014, CoRR abs/1408.
[80] F. Prost, H. Qian, Q. Chen, E.H. Chi, J. Chen, A. Beutel, Toward a better trade-
6491.
off between performance and fairness with kernel-based distribution matching,
[50] B. Imana, A. Korolova, J.S. Heidemann, Auditing for discrimination in
2019, CoRR abs/1910.11779.
algorithms delivering job ads, in: WWW, ACM / IW3C2, 2021, pp. 3767–3778.
[81] A. Beutel, J. Chen, T. Doshi, H. Qian, L. Wei, Y. Wu, L. Heldt, Z. Zhao, L. Hong,
[51] V. Do, S. Corbett-Davies, J. Atif, N. Usunier, Online certification of preference-
E.H. Chi, C. Goodrow, Fairness in recommendation ranking through pairwise
based fairness for personalized recommender systems, in: Proceedings of the
comparisons, in: KDD, ACM, 2019, pp. 2212–2220.
AAAI Conference on Artificial Intelligence, 2022, pp. 6532–6540.
[52] Y. Zhao, W. Ma, Y. Jiang, J. Zhan, A moocs recommender system based on [82] M. Wang, X. Zheng, Y. Yang, K. Zhang, Collaborative filtering with social
user’s knowledge background, in: KSEM, Springer, 2021, pp. 140–153. exposure: A modular approach to social recommendation, in: AAAI, AAAI Press,
[53] Z. Zheng, Z. Qiu, T. Xu, X. Wu, X. Zhao, E. Chen, H. Xiong, Cbr: Context bias 2018, pp. 2516–2523.
aware recommendation for debiasing user modeling and click prediction, in: [83] J. Wasilewski, N. Hurley, Incorporating diversity in a learning to rank
Proceedings of the ACM Web Conference 2022, 2022, pp. 2268–2276. recommender system, in: FLAIRS Conference, AAAI Press, 2016, pp. 572–578.
[54] G. Jawaheer, M. Szomszor, P. Kostkova, Comparison of implicit and explicit [84] D. Kiswanto, D. Nurjanah, R. Rismala, Fairness aware regularization on
feedback from an online music recommendation service, in: HetRec@RecSys, a learning-to-rank recommender system for controlling popularity bias in
ACM, 2010, pp. 47–51. e-commerce domain, in: 2018 International Conference on Information
[55] F. Liang, W. Pan, Z. Ming, Fedrec++: Lossless federated recommendation with Technology Systems and Innovation, ICITSI, IEEE, 2018, pp. 16–21.
explicit feedback, in: AAAI, AAAI Press, 2021, pp. 4224–4231. [85] H. Abdollahpouri, R. Burke, B. Mobasher, Controlling popularity bias in
[56] X. Zhang, S. Dai, J. Xu, Z. Dong, Q. Dai, J.R. Wen, Counteracting user learning-to-rank recommendation, in: RecSys, ACM, 2017, pp. 42–46.
attention bias in music streaming recommendation via reward modification, [86] Y. Hu, Y. Koren, C. Volinsky, Collaborative filtering for implicit feedback
in: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery datasets, in: ICDM, IEEE Computer Society, 2008, pp. 263–272.
and Data Mining, 2022, pp. 2504–2514. [87] B. Rastegarpanah, K.P. Gummadi, M. Crovella, Fighting fire with fire: Using
[57] S. Gupta, H. Wang, Z.C. Lipton, Y. Wang, Correcting exposure bias for link antidote data to improve polarization and fairness of recommender systems, in:
recommendation, in: ICML, PMLR, 2021, pp. 3953–3963. WSDM, ACM, 2019, pp. 231–239.
[58] M. Cheng, F. Yuan, Q. Liu, S. Ge, Z. Li, R. Yu, D. Lian, S. Yuan, E.
[88] R. Burke, N. Sonboli, A. Ordonez-Gauger, Balanced neighborhoods for
Chen, Learning recommender systems with implicit feedback via soft target
multi-sided fairness in recommendation, in: FAT, PMLR, 2018, pp. 202–214.
enhancement, in: SIGIR, ACM., 2021, pp. 575–584.
[89] B. Edizel, F. Bonchi, S. Hajian, A. Panisson, T. Tassa, Fairecsys: mitigating
[59] R.C. Chen, Q. Ai, G. Jayasinghe, W.B. Croft, Correcting for recency bias in job
algorithmic bias in recommender systems, Int. J. Data Sci. Anal. 9 (2020)
recommendation, 2019, pp. 2185–2188.
197–213.
[60] D. Cohen, M. Aharon, Y. Koren, O. Somekh, R. Nissim, Expediting exploration
by attribute-to-feature mapping for cold-start recommendations, in: RecSys, [90] R. Borges, K. Stefanidis, F2VAE: a framework for mitigating user unfairness in
ACM, 2017, pp. 184–192. recommendation systems, in: SAC, ACM, 2022, pp. 1391–1398.
[61] M. Vartak, A. Thiagarajan, C. Miranda, J. Bratman, H. Larochelle, A meta- [91] C. Wu, F. Wu, X. Wang, Y. Huang, X. Xie, Fairness-aware news recommenda-
learning perspective on cold-start recommendations for items, in: Advances in tion with decomposed adversarial learning, in: AAAI, AAAI Press, 2021, pp.
Neural Information Processing Systems, Vol. 30, 2017. 4462–4469.
[62] Z. Zhu, J. Kim, T. Nguyen, A. Fenton, J. Caverlee, Fairness among new items [92] L. Boratto, G. Fenu, M. Marras, Interplay between upsampling and regulariza-
in cold start recommender systems, in: SIGIR, ACM, 2021, pp. 767–776. tion for provider fairness in recommender systems, User Model. User Adapt.
[63] Q. Wan, X. He, X. Wang, J. Wu, W. Guo, R. Tang, Cross pairwise ranking for Interact. 31 (2021) 421–455.
unbiased item recommendation, in: Proceedings of the ACM Web Conference [93] Z. Zhu, X. Hu, J. Caverlee, Fairness-aware tensor-based recommendation, in:
2022, 2022, pp. 2370–2378. CIKM, ACM, 2018, pp. 1153–1162.

19
D. Jin et al. Information Fusion 100 (2023) 101906

[94] W. Sun, S. Khenissi, O. Nasraoui, P. Shafto, Debiasing the human-recommender [119] C. Dickens, R. Singh, L. Getoor, Hyperfair: A soft approach to integrating
system feedback loop in collaborative filtering, in: WWW (Companion Volume), fairness criteria, 2020, CoRR abs/2009.08952.
ACM, 2019, pp. 645–651. [120] R. Mehrotra, J. McInerney, H. Bouchard, M. Lalmas, F. Diaz, Towards a fair
[95] Z. Qin, S.J. Chen, D. Metzler, Y. Noh, J. Qin, X. Wang, Attribute-based marketplace: Counterfactual evaluation of the trade-off between relevance,
propensity for unbiased learning in recommender systems: Algorithm and case fairness & satisfaction in recommendation systems, in: CIKM, ACM, 2018, pp.
studies, 2020, pp. 2359–2367. 2243–2251.
[96] M. Zhao, L. Wu, Y. Liang, L. Chen, J. Zhang, Q. Deng, K. Wang, X. Shen, [121] A. Singh, T. Joachims, Fairness of exposure in rankings, KDD (2018)
T. Lv, R. Wu, Investigating accuracy-novelty performance for graph-based 2219–2228.
collaborative filtering, in: SIGIR, ACM, 2022, pp. 50–59. [122] K. Yang, J. Stoyanovich, Measuring fairness in ranked outputs, in: SSDBM, ACM,
[97] Y. Zhang, F. Feng, X. He, T. Wei, C. Song, G. Ling, Y. Zhang, Causal intervention 2017, pp. 221–226.
for leveraging popularity bias in recommendation, in: Proceedings of the [123] S.C. Geyik, S. Ambler, K. Kenthapadi, Fairness-aware ranking in search &
44th International ACM SIGIR Conference on Research and Development in recommendation systems with application to linkedin talent search, 2019, CoRR
Information Retrieval, 2021, pp. 11–20. abs/1905.01989.
[98] T. Wei, F. Feng, J. Chen, Z. Wu, J. Yi, X. He, Model-agnostic counterfactual rea- [124] M. Tavakol, Fair classification with counterfactual learning, in: SIGIR, ACM,
soning for eliminating popularity bias in recommender system, in: Proceedings 2020, pp. 2073–2076.
of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, [125] N. Akpinar, C. DiCiccio, P. Nandy, K. Basu, Long-term dynamics of fairness
2021, pp. 1791–1800. intervention in connection recommender systems, in: AIES, ACM, 2022, pp.
[99] J. Huang, H. Oosterhuis, M. de Rijke, It is different when items are older: 22–35.
Debiasing recommendations when selection bias and user preferences are [126] M. Mladenov, E. Creager, K. Ben-Porat, R.S. Zemel, C. Boutilier, Optimizing
dynamic, in: Proceedings of the Fifteenth ACM International Conference on long-term social welfare in recommender systems: A constrained matching
Web Search and Data Mining, 2022, pp. 381–389. approach, in: ICML, PMLR, 2020, pp. 6987–6998.
[100] R. Acharyya, S. Das, A. Chattoraj, M.I. Tanveer, Fairyted: A fair rating [127] V. Do, S. Corbett-Davies, J. Atif, N. Usunier, Online certification of preference-
predictor for ted talk data, in: Proceedings of the AAAI Conference on Artificial based fairness for personalized recommender systems, in: AAAI, AAAI Press,
Intelligence, 2020, pp. 338–345. 2022, pp. 6532–6540.
[101] X. Wang, R. Zhang, Y. Sun, J. Qi, Combating selection biases in recommender [128] E.H. Simpson, The interpretation of interaction in contingency tables, J. R. Stat.
systems with a few unbiased ratings, in: WSDM, ACM, 2021, pp. 427–435. Soc. Ser. B Stat. Methodol. 13 (1951) 238–241.
[102] L. Xia, C. Huang, C. Zhang, Self-supervised hypergraph transformer for [129] T. Xiao, S. Wang, Towards unbiased and robust causal ranking for recommender
recommender systems, in: KDD, ACM, 2022, pp. 2100–2109. systems, in: Proceedings of the Fifteenth ACM International Conference on Web
[103] T. Schnabel, A. Swaminathan, A. Singh, N. Chandak, T. Joachims, Recom- Search and Data Mining, 2022, pp. 1158–1167.
mendations as treatments: Debiasing learning and evaluation, in: International [130] Z. Wang, S. Shen, Z. Wang, B. Chen, X. Chen, J.R. Wen, Unbiased sequential
Conference on Machine Learning, PMLR, 2016, pp. 1670–1679. recommendation with latent confounders, in: Proceedings of the ACM Web
[104] Y. Li, H. Chen, S. Xu, Y. Ge, Y. Zhang, Towards personalized fairness based on Conference 2022, 2022, pp. 2195–2204.
causal notion, in: Proceedings of the 44th International ACM SIGIR Conference
[131] W. Wang, F. Feng, X. He, X. Wang, T. Chua, Deconfounded recommendation
on Research and Development in Information Retrieval, 2021, pp. 1054–1063.
for alleviating bias amplification, 2021, CoRR abs/2105.10648.
[105] X. Qian, Y. Xu, F. Lv, S. Zhang, Z. Jiang, Q. Liu, X. Zeng, T. Chua, F. Wu,
[132] Y. Ge, J. Tan, Y. Zhu, Y. Xia, J. Luo, S. Liu, Z. Fu, S. Geng, Z. Li, Y. Zhang,
Intelligent request strategy design in recommender system, in: KDD, ACM, 2022,
Explainable fairness in recommendation, in: E. Amigó, P. Castells, J. Gonzalo,
pp. 3772–3782.
B. Carterette, J.S. Culpepper, G. Kazai (Eds.), SIGIR ’22: The 45th International
[106] Z. Wang, Y. He, J. Liu, W. Zou, P.S. Yu, P. Cui, Invariant preference learning for
ACM SIGIR Conference on Research and Development in Information Retrieval,
general debiasing in recommendation, in: Proceedings of the 28th ACM SIGKDD
Madrid, Spain, July 11-15 2022, ACM, 2022, pp. 681–691, http://dx.doi.org/
Conference on Knowledge Discovery and Data Mining, 2022, pp. 1969–1978.
10.1145/3477495.3531973.
[107] S. Rajanala, A. Pal, M. Singh, R.C.W. Phan, K. Wong, Descover: Debiased
[133] H. Liu, Y. Wang, H. Lin, B. Xu, N. Zhao, Mitigating sensitive data exposure with
semantic context prior for venue recommendation, in: Proceedings of the
adversarial learning for fairness recommendation systems, Neural Comput. Appl.
45th International ACM SIGIR Conference on Research and Development in
(2022) 1–15.
Information Retrieval, 2022, pp. 2456–2461.
[134] J. Li, Y. Ren, K. Deng, Fairgan: Gans-based fairness-aware learning for recom-
[108] H. Liu, N. Zhao, X. Zhang, H. Lin, L. Yang, B. Xu, Y. Lin, W. Fan, Dual
mendations with implicit feedback, in: Proceedings of the ACM Web Conference
constraints and adversarial learning for fair recommenders, Knowl.-Based Syst.
2022, 2022, pp. 297–307.
239 (2022) 108058.
[135] Y. Li, H. Chen, S. Xu, Y. Ge, J. Tan, S. Liu, Y. Zhang, Fairness in
[109] C. Rus, J. Luppes, H. Oosterhuis, G.H. Schoenmacker, Closing the gender wage
recommendation: A survey, 2022, CoRR abs/2205.13619.
gap: Adversarial fairness in job recommendation, 2022, CoRR abs/2209.09592.
[136] W. Liu, F. Liu, R. Tang, B. Liao, G. Chen, P.A. Heng, Balancing accuracy
[110] Y. Ge, X. Zhao, L. Yu, S. Paul, D. Hu, C.C. Hsieh, Y. Zhang, Toward pareto
and fairness for interactive recommendation with reinforcement learning, 2021,
efficient fairness-utility trade-off in recommendation through reinforcement
arXiv preprint arXiv:2106.13386.
learning, in: Proceedings of the Fifteenth ACM International Conference on Web
[137] Z. Deng, L. Huang, C. Wang, J. Lai, P.S. Yu, Deepcf: A unified framework of
Search and Data Mining, 2022, pp. 316–324.
representation learning and matching function learning in recommender system,
[111] Z. Fu, Y. Xian, S. Geng, G. de Melo, Y. Zhang, Popcorn: Human-in-the-loop
in: AAAI, AAAI Press, 2019, pp. 61–68.
popularity debiasing in conversational recommender systems, in: CIKM, 2021,
[138] S. Rendle, C. Freudenthaler, Z. Gantner, L. Schmidt-Thieme, BPR: bayesian
pp. 494–503, http://dx.doi.org/10.1145/3459637.3482461.
personalized ranking from implicit feedback, in: UAI, AUAI Press, 2009, pp.
[112] H. Liu, D. Tang, J. Yang, X. Zhao, H. Liu, J. Tang, Y. Cheng, Rating distribution
452–461.
calibration for selection bias mitigation in recommendations, in: Proceedings of
the ACM Web Conference 2022, 2022, pp. 2048–2057. [139] Z. Fu, Y. Xian, R. Gao, J. Zhao, Q. Huang, Y. Ge, S. Xu, S. Geng, C. Shah,
Y. Zhang, et al., Fairness-aware explainable recommendation over knowledge
[113] C. Zhou, J. Ma, J. Zhang, J. Zhou, H. Yang, Contrastive learning for debiased
graphs, in: Proceedings of the 43rd International ACM SIGIR Conference on
candidate generation in large-scale recommender systems, in: Proceedings of
Research and Development in Information Retrieval, 2020, pp. 69–78.
the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining,
2021, pp. 3985–3995. [140] G.K. Patro, A. Biswas, N. Ganguly, K.P. Gummadi, A. Chakraborty, Fairrec:
[114] Q. Shen, W. Tao, J. Zhang, H. Wen, Z. Chen, Q. Lu, Sar-net: A scenario-aware Two-sided fairness for personalized recommendations in two-sided platforms,
ranking network for personalized fair recommendation in hundreds of travel in: WWW, ACM / IW3C2, 2020, pp. 1194–1204.
scenarios, 2021, CoRR abs/2110.06475. [141] V.W. Anelli, T. Di Noia, F.A. Merra, The idiosyncratic effects of adversarial
[115] C.T. Li, C. Hsu, Y. Zhang, Fairsr: Fairness-aware sequential recommenda- training on bias in personalized recommendation learning, in: Fifteenth ACM
tion through multi-task learning with preference graph embeddings, in: ACM Conference on Recommender Systems, 2021, pp. 730–735.
Transactions on Intelligent Systems and Technology, TIST, 2022, pp. 131–121. [142] Z. Zhu, J. Caverlee, Fighting mainstream bias in recommender systems via local
[116] M. Naghiaei, H.A. Rahmani, Y. Deldjoo, Cpfair: Personalized consumer and fine tuning, in: Proceedings of the Fifteenth ACM International Conference on
producer fairness re-ranking for recommender systems, 2022, arXiv preprint Web Search and Data Mining, 2022, pp. 1497–1506.
arXiv:2204.08085. [143] Z. Zhu, Y. He, X. Zhao, Y. Zhang, J. Wang, J. Caverlee, Popularity-opportunity
[117] M. Mansoury, H. Abdollahpouri, M. Pechenizkiy, B. Mobasher, R. Burke, bias in collaborative filtering, in: Proceedings of the 14th ACM International
Fairmatch: A graph-based approach for improving aggregate diversity in recom- Conference on Web Search and Data Mining, 2021, pp. 85–93.
mender systems, in: Proceedings of the 28th ACM Conference on User Modeling, [144] R. Sato, Enumerating fair packages for group recommendations, in: Proceedings
Adaptation and Personalization, 2020, pp. 154–162. of the Fifteenth ACM International Conference on Web Search and Data Mining,
[118] Y. Wu, J. Cao, G. Xu, Y. Tan, TFROM: A two-sided fairness-aware recom- 2022, pp. 870–878.
mendation model for both customers and providers, 2021, CoRR abs/2104. [145] H. Wu, B. Mitra, C. Ma, F. Diaz, X. Liu, Joint multisided exposure fairness for
09024. recommendation, 2022, arXiv preprint arXiv:2205.00048.

20
D. Jin et al. Information Fusion 100 (2023) 101906

[146] C. Zhou, J. Ma, J. Zhang, J. Zhou, H. Yang, Contrastive learning for debiased [173] A.B. Melchiorre, E. Zangerle, M. Schedl, Personality bias of music recommen-
candidate generation at scale, 2020, CoRR abs/2005.12964. dation algorithms, in: Fourteenth ACM Conference on Recommender Systems,
[147] E. Mena-Maldonado, P. Cañamares, Y. Ren, M. Sanderson, Popularity bias in 2020, pp. 533–538.
false-positive metrics for recommender systems evaluation, ACM Trans. Inf. [174] S. Xu, Y. Li, S. Liu, Z. Fu, Y. Zhang, Learning post-hoc causal explanations for
Syst. 39 (2021) 1–43. recommendation, 2020, CoRR abs/2006.16977.
[148] H. Abdollahpouri, M. Mansoury, R. Burke, B. Mobasher, The connection [175] G. Cornacchia, F. Narducci, A. Ragone, A general model for fair
between popularity bias, calibration, and fairness in recommendation, in: and explainable recommendation in the loan domain (short paper), in:
Fourteenth ACM Conference on Recommender Systems, 2020, pp. 726–731. KaRS/ComplexRec@RecSys, CEUR-WS.org, 2021.
[149] R. Islam, K.N. Keya, Z. Zeng, S. Pan, J. Foulds, Debiasing career recommen-
[176] P. Dokoupil, L. Peska, Robustness against polarity bias in decoupled group rec-
dations with neural fair collaborative filtering, in: Proceedings of the Web
ommendations evaluation, in: Adjunct Proceedings of the 30th ACM Conference
Conference 2021, 2021, pp. 3779–3790.
on User Modeling, Adaptation and Personalization, 2022, pp. 302–307.
[150] D. Liu, P. Cheng, H. Zhu, Z. Dong, X. He, W. Pan, Z. Ming, Mitigating
[177] Z. Zhu, S. Si, J. Wang, J. Xiao, Cali3f: Calibrated fast fair federated
confounding bias in recommendation via information bottleneck, in: Fifteenth
recommendation system, 2022, arXiv preprint arXiv:2205.13121.
ACM Conference on Recommender Systems, 2021, pp. 351–360.
[151] Y. Saito, S. Yaginuma, Y. Nishino, H. Sakata, K. Nakata, Unbiased recommender [178] D. Sacharidis, C.P. Mukamakuza, H. Werthner, Fairness and diversity in social-
learning from missing-not-at-random implicit feedback, in: Proceedings of the based recommender systems, in: UMAP (Adjunct Publication), ACM, 2020, pp.
13th International Conference on Web Search and Data Mining, 2020, pp. 83–88.
501–509. [179] M. Mansoury, H. Abdollahpouri, J. Smith, A. Dehpanah, M. Pechenizkiy, B.
[152] J. Huang, H. Oosterhuis, M. de Rijke, H. van Hoof, Keeping dataset biases out Mobasher, Investigating potential factors associated with gender discrimination
of the simulation: A debiased simulator for reinforcement learning based rec- in collaborative recommender systems, in: FLAIRS Conference, AAAI Press,
ommender systems, in: Fourteenth ACM Conference on Recommender Systems, 2020, pp. 193–196.
2020, pp. 190–199. [180] J. Leonhardt, A. Anand, M. Khosla, User fairness in recommender systems, in:
[153] C. Yang, Q. Wu, J. Jin, X. Gao, J. Pan, G. Chen, Trading hard negatives and true WWW (Companion Volume), ACM, 2018, pp. 101–102.
negatives: A debiased contrastive collaborative filtering approach, in: IJCAI, [181] L. Schelenz, Diversity-aware recommendations for social justice? exploring user
ijcai.org, 2022, pp. 2355–2361. diversity and fairness in recommender systems, in: UMAP (Adjunct Publication),
[154] C. Yang, Q. Wu, J. Jin, X. Gao, J. Pan, G. Chen, Trading hard negatives and ACM, 2021, pp. 404–410.
true negatives: A debiased contrastive collaborative filtering approach, 2022,
arXiv preprint arXiv:2204.11752.
[155] T. Qi, F. Wu, C. Wu, P. Sun, L. Wu, X. Wang, Y. Huang, X. Xie, Profairrec:
Provider fairness-aware news recommendation, 2022, arXiv preprint arXiv:
2204.04724.
[156] Y. Sun, J. Pan, A. Zhang, A. Flores, FM2: field-matrixed factorization machines
Di Jin received the Ph.D. degree in computer science from
for recommender systems, in: WWW, ACM/ IW3C2, 2021, pp. 2828–2837.
Jilin University, Changchun, China, in 2012. He was a
[157] D. Antognini, B. Faltings, Fast multi-step critiquing for vae-based recommender
research scholar in DMG at UIUC from 2019 to 2020. He
systems, in: RecSys, ACM, 2021, pp. 209–219.
is currently a Professor at the College of Intelligence and
[158] J. Zou, E. Kanoulas, P. Ren, Z. Ren, A. Sun, C. Long, Improving conversational
Computing, Tianjin University, Tianjin, China. His research
recommender systems via transformer-based sequential modelling, in: SIGIR,
interests include graph data mining and graph machine
ACM, 2022, pp. 2319–2324.
learning, especially on community detection, network em-
[159] W. Krichene, S. Rendle, On sampled metrics for item recommendation, bedding, and GNNs. To date, he has published more than
Commun. ACM 65 (2022) 75–83. 100 research papers in top-tier journals and conferences,
[160] Y. Wang, L. Wang, Y. Li, D. He, T. Liu, A theoretical analysis of NDCG type including the TKDE, TNNLS, TYCB, AAAI, IJCAI, NeurIPS,
ranking measures, in: COLT, JMLR.org, 2013, pp. 25–54. and WWW. He was the recipient of the Best Paper Award
[161] D.C. da Silva, M.G. Manzato, F.A. Durão, Exploiting personalized calibration and Runner-up of WWW 2021, Best Student Paper Award
metrics for fairness recommendation, Expert Syst. Appl. 181 (2021) 115112. Runner-up of ICDM 2021, and Rising Star Award of ACM
[162] X. Wu, H. Chen, J. Zhao, L. He, D. Yin, Y. Chang, Unbiased learning to Tianjin in 2018.
rank in feeds recommendation, in: Proceedings of the 14th ACM International
Conference on Web Search and Data Mining, 2021, pp. 490–498.
[163] Gómez E., C.S. Zhang, L. Boratto, M. Salamó, G. Ramos, Enabling cross-
continent provider fairness in educational recommender systems, Future Gener.
Comput. Syst. 127 (2022) 435–447.
[164] M. Marras, L. Boratto, G. Ramos, G. Fenu, Equality of learning opportunity via Luzhi Wang is currently a third-year Ph.D. student at
individual fairness in personalized recommendations, Int. J. Artif. Intell. Educ. Tianjin University, under the supervision of Prof. Di Jin
(2021) 1–49. and Prof. Shirui Pan. Her primary research interests re-
[165] L. Boratto, G. Fenu, M. Marras, The effect of algorithmic bias on recom- volve around graph representation learning, self-supervised
mender systems for massive open online courses, in: European Conference on learning, and graph neural networks. She has published
Information Retrieval, Springer, 2019, pp. 457–472. numerous papers in top-tier venues and journals, including
[166] T. Qi, F. Wu, C. Wu, Y. Huang, Pp-rec: News recommendation with personalized the IJCAI, WWW, and TWEB, etc.
user interest and time-aware news popularity, 2021, arXiv preprint arXiv:
2106.01300.
[167] B. Kirdemir, J. Kready, E. Mead, M.N. Hussain, N. Agarwal, D. Adjeroh,
Assessing bias in youtube’s video recommendation algorithm in a cross-lingual
and cross-topical context, in: International Conference on Social Computing,
Behavioral-Cultural Modeling and Prediction and Behavior Representation in
Modeling and Simulation, Springer, 2021, pp. 71–80.
[168] R. Zhan, C. Pei, Q. Su, J. Wen, X. Wang, G. Mu, D. Zheng, P. Jiang, K. Gai, De- He Zhang is currently a Ph.D. student with the Faculty
confounding duration bias in watch-time prediction for video recommendation, of Information Technology, Monash University, Australia.
in: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery His research interests include graph neural networks, graph
and Data Mining, 2022, pp. 4472–4481. attack and defence, and trustworthy graph learning.
[169] C. Wu, F. Wu, T. Qi, Y. Huang, Fairrank: Fairness-aware single-tower ranking
framework for news recommendation, 2022, arXiv preprint arXiv:2204.00541.
[170] A. Brek, Z. Boufaïda, Semantic approaches survey for job recommender systems,
in: RIF, CEUR-WS.org, 2022, pp. 101–111.
[171] J. Yi, F. Wu, C. Wu, Q. Li, G. Sun, X. Xie, Debiasedrec: Bias-aware user
modeling and click prediction for personalized news recommendation, 2021,
arXiv preprint arXiv:2104.07360.
[172] D. Shakespeare, L. Porcaro, C. Gómez, Exploring artist gender bias in music
recommendation, 2020, arXiv preprint arXiv:2009.01715.

21
D. Jin et al. Information Fusion 100 (2023) 101906

Yizhen Zheng is a second-year Ph.D. student at Monash Dr. Feng Xia is a Professor in School of Computing Tech-
University, supervised by A/Prof. Vincent CS Lee and Prof. nologies, RMIT University, Australia. He is/was on the
Shirui Pan. His research interests lie in graph representation Editorial Boards of over 10 int’l journals. He has served as
learning, self-supervised learning and graph neural net- the General Chair, PC Chair, Workshop Chair, or Publicity
works. He received the M.Bus.Info.Sys degree from Monash Chair of over 30 int’l conferences and workshops, and PC
University as a Top 1 student in 2020, M.Bus degree from Member of over 90 conferences. Dr. Xia has authored/co-
Monash University in 2018 and B.B.A degree from Shenzhen authored two books and over 300 scientific papers in
University in 2016. He published papers on top venues int’l journals and conferences (such as IEEE TAI, TKDE,
and journals such as ICML, NeurIPS, WWW, AAAI, TNNLS, TNNLS, TC, TMC, TPDS, TBD, TCSS, TNSE, TETCI, TETC,
ICDM, IJCAI, etc. He is invited to review conferences and THMS, TVT, TITS, TASE, ACM TKDD, TIST, TWEB, TOMM,
journals including TPAMI, TNNLS, ECAI, ICDM, TWEB, PR, WWW, AAAI, SIGIR, WSDM, CIKM, JCDL, EMNLP, and
etc. INFOCOM). He was recognized as a Highly Cited Researcher
(2019). Dr. Xia received a number of prestigious awards,
including IEEE DSS 2021 Best Paper Award, IEEE Vehicular
Weiping Ding (M’16-SM’19) received the Ph.D. degree in
Technology Society 2020 Best Land Transportation Paper
Computer Science, Nanjing University of Aeronautics and
Award, ACM/IEEE JCDL 2020 The Vannevar Bush Best Pa-
Astronautics, Nanjing, China, in 2013. In 2016, He was
per Honorable Mention, IEEE CSDE 2020 Best Paper Award,
a Visiting Scholar at National University of Singapore,
WWW 2017 Best Demo Award, IEEE DataCom 2017 Best
Singapore. From 2017 to 2018, he was a Visiting Pro-
Paper Award, IEEE UIC 2013 Best Paper Award, and IEEE
fessor at University of Technology Sydney, Australia. He
Access Outstanding Associate Editor. He has been invited as
is a Full Professor with the School of Information Sci-
Keynote Speaker at seven int’l conferences, and delivered
ence and Technology, Nantong University, Nantong, China.
a number of Invited Talks at int’l conferences and many
His main research directions involve deep neural net-
universities worldwide. His research interests include data
works, multimodal machine learning, and medical images
science, artificial intelligence, graph learning, and systems
analysis. He has published over 200 articles, including
engineering. He is a Senior Member of IEEE and ACM, and
over 90 IEEE Transactions papers. His fifteen authored/co-
an ACM Distinguished Speaker.
authored papers have been selected as ESI Highly Cited
Papers. He serves as an Associate Editor/Editorial Board
member of IEEE Transactions on Neural Networks and Shirui Pan received a Ph.D. in computer science from
Learning Systems, IEEE Transactions on Fuzzy Systems, the University of Technology Sydney (UTS), Ultimo, NSW,
IEEE/CAA Journal of Automatica Sinica, IEEE Transactions Australia. He is an ARC Future Fellow and a professor
on Intelligent Transportation Systems, IEEE Transactions at the School of Information and Communication Tech-
on Intelligent Vehicles, IEEE Transactions on Emerging nology, Griffith University, Australia. His research interests
Topics in Computational Intelligence, IEEE Transactions include data mining and machine learning. To date, Dr
on Artificial Intelligence, Information Fusion, Information Pan has published over 150 research papers in top-tier
Sciences, Neurocomputing, Applied Soft Computing. He is journals and conferences, including the IEEE Transactions
the Leading Guest Editor of Special Issues in several presti- on Pattern Analysis and Machine Intelligence (TPAMI), IEEE
gious journals, including IEEE Transactions on Evolutionary Transactions on Knowledge and Data Engineering (TKDE),
Computation, IEEE Transactions on Fuzzy Systems, and IEEE Transactions on Neural Networks and Learning Systems
Information Fusion. (TNNLS), ICML, NeurIPS, KDD, AAAI, IJCAI, WWW, CVPR,
and ICDM. His paper received the Best Student Paper Award
of IEEE ICDM 2020. He is recognized as one of the AI 2000
AAAI/IJCAI Most Influential Scholars in Australia (2022,
2021).

22