See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/344197796

The use of Recommender Systems in web technology and an in-depth analysis

of Cold State problem

Preprint · September 2020

CITATIONS READS

0 230

2 authors:

Denis Selimi Krenare Pireva Nuci

University for Business and Technology University for Business and Technology
1 PUBLICATION 0 CITATIONS 37 PUBLICATIONS 106 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Design of Kosovo Current Research Information System - the national repository with respect to data for research and research data View project

Erasmus KA2 View project

All content following this page was uploaded by Krenare Pireva Nuci on 28 October 2020.

The user has requested enhancement of the downloaded file.

 The use of Recommender Systems in web technology and an in-depth
analysis of Cold State problem

Denis Selimi1, Krenare PIREVA NUCI2

12
University for Business and Technology, Faculty of Computer Science and Engineering

[email protected], [email protected]

Abstract. In the WWW (World Wide Web), dynamic development and spread of data has resulted a
tremendous amount of information available on the Internet, yet user is unable to find relevant information
in a short span of time. Consequently, a system called recommendation system developed to help users find
their infromation with ease through their browsing activities. In other words, recommender systems are
tools for interacting with large amount of information that provide personalized view for prioritizing items
likely to be of keen for users. They have developed over the years in artificial intelligence techniques that
include machine learning and data mining amongst many to mention. Furthermore, the recommendation
systems have personalized on an e-commerce, on-line applications such as Amazon.com, Netflix, and
Booking.com. As a result, this has inspired many researchers to extend the reach of recommendation
systems into new sets of challenges and problem areas that are yet to be truly solved, primarily a problem
with the case of making a recommendation to a new user that is called cold-state (i.e. cold-start) user
problem where the new user might likely not yield much of information searched. Therfore, the purpose of
this paper is to tackle the said cold-start problem with a few effecient methods and challenges, as well as
identify and overview the current state of recommendation system as a whole.

Keywords: Recommender Systems, Cold State, Collaborative filtering, Content filtering

1. Introduction
In this ever-changing technological world of advancement with diverse amount of data on
the internet, has led the user to be overloaded with information [1]. The recommender system
(RS) thus allows us to manage and effectively deal with said overload, by classifying the set
of items into categorized niches that later can be recommended by users who search for it,
whether be it textual, audio, image, and video in which are constantly updated.
Recommendation system nowadays, allows us to explore variety of items that may make user
content from recommendations they provide – be it in the movies, music, or news. Such
example amongst many, are Amazon that recommends books, Booking.com recommending
accommodations for the users, Ebay known for using item-to-item based collaboration, and
Netflix who recommends movies - all that provide and benefit from recommending to their
clients [2]. This recommendation is determined by variety techniques, as highlighted by
many scientific studies [3]. Giving examples by some of the techniques, include an
improvement over the transaction costs of finding and selecting specific items in online
shopping environment, for e-commerce purposes for selling more products as a tool that

1
searches users’ preferences, and in scientific libraries whereas the RS aids in users to search
beyond the specified niches. Therefore, it is indispensable to understand the concept of
converting the data information into recommendation techniques within a system that will be
a convenient approach for the users to not be overloaded.
The combination of RS concepts and using its techniques are a new understanding for
creating a recommendation e-commerce site to predict the users’ preferences on an item. The
work of this topic has been achieved using data from various sources that have been critically
checked and analyzed.
In this paper, a new framework based on web applications (i.e. e-commerce sites) of
recommendation systems and its techniques are proposed to overcome the drawbacks
associated mostly from the cold state situation that will further be defined in the problem
declaration section within paper. Although the main point within this paper relies on its
problem with cold state in problem declaration section, in the section two lists important the
concepts of recommendation system in which it uses specific variation of algorithms and
techniques, mostly being the content-based filtering and its user-based analysis, collaborative
filtering technique. In the next section are included theories, terms, and substantive findings
in which will help tremendously in analyzing the information concepts, i.e. content-based
filtering, and by giving examples of other users’ accesses of behaviors such as collaborative
filtering.

1.1. Problem declaration

The use of recommender system in intelligent application is common nowadays; however,
besides the advantages that are coming with the new technologies, there are still gaps that are
phased in the very beginning of these systems. One of the main gaps that the system may
face is the cold state situation, where the system is new and there is no data recorded
regarding the users’ profiles. Thus, the main problem of this paper is how we can escape this
situation and whether the system will be able to have the intelligent part from the very
beginning.
The aim of the problem declaration within recommendation system is to tackle the cold-state
(i.e. cold-start) situation in web applications that tends to halt certainty in e-commerce
companies or sites (e.g. Amazon, Netflix) and/or aforementioned Booking.com site. So in
order to give possible treatments, in the section 5.0 will be further stated in ways to directly
solidify the cold-state situation itself. There are also mentionings of techniques in the
previous section that aid to define the RS.
The objectives of this paper is to: (1): Review the use of recommender systems in intelligent
systems, (2) to list and compare a set of algorithms used in RS and (3) to list a few scenarios
to possibly treat in continuous cold state situation from certain users.
The paper is strucutured as follows, section 2 gives the general knowledge how recommender
systems work and the related work in this domain, continuing with section 3, which give an
in-depth analysis of Cold State problem within initial phase of Recommender Systems.The
discussion is provided in section 4, and finally concluding the paper in section 5.

2
2. General Concepts of Recommender Systems and Related work
In this section is addressed the use of algorithms and different style of techniques that help
identify recommendation system itself, by initially listing the algorithms used in
recommendation systems and its data mining techniques that identifies a good algorithm for
RS (see Figure 1).

Recommender System

Content-based filtering Collaborative filtering Hybrid filtering

technique technique technique

Model-based filtering Memory-based filtering

technique technique

Clustering
Regression
User-based Item-based
Matrix completion technique

Figure 1: Block diagram of recommendation system techniques.

As shown in Figure 1 Recommender System use Content, Collaborative and Hybrid

Filtering techniques to recommend a number of items depending in which domain we are
applying them. Collaborative filtering technique will specify its in-depth relations between

3
the user and set of items, giving us two subsections: Memory-based technique and mode-
based technique within collaborative filtering. Once the filtering techniques are mentioned,
the hybrid filtering has its importance to tie and conclude how the relations between CF and
CBF can be combined to yield a better recommendation system.
2.1 The algorithms used in RS
The amount of data being generated nowadays has increased exponentially, whether being
business related such as purchase transactions, to scientific domains such as satellite image
datas. Nowadays, the algorithms within RS are used with data mining, in which it exhibits
the following characteristics:
High-level efficiency: The information that it is discovered should in a high-level language
which is understood by human users. Furthermore, the discovery process of said information
is efficient and yields the running times for the mass sizes of databases acceptable.
Unique results with accuracy: Said discovered information should have an interesting,
unique result according to user biases. In particular, having a unique result with accurate
contents of databases portrays that patterns are novel and perhaps useful.
With that said, the data mining is crucial when it comes to recommendation system and data
science branch itself, as it helps in research fields such as database management, statistics,
and machine learning; therefore, there are three classes of data mining algorithms that are
applied in various niches of applications, and they are as follows:
Classification: This type of algorithm is widely known amongst data mining algorithm and
has been an ongoing study in the research field of machine learning for several decades. The
main goal of classification algorithm is to classify cases into different classes that gets based
on attributes among a set of objects within database. Such example in classification gets used
in medical diagnosis, target niches of marketing, fraud detection, as well as the information
retrieval.
Association rule mining: This is a method – or a procedure – that is meant to find frequent
patterns. The output of this method has a set of rules where the occurrence of items in
precondition implies the co-occurrence of items in precondition. They are found in various
kinds of databases such as relational databases, transactions, and e-commerce sites.
Clustering: The clustering algorithm, or in other words called unsupervised classification,
is the process of grouping physical or abstract objects in such a way, where the objects in the
same group are more similar to each other than to those other groups. Hencemore, clustering
analysis helps constructing a meaningful partitioning on the mass size of objects based on
the characteristics. The clusting is used and will be mentioned more in the model-based
collaborative technique within recommendation system. Below is a table used as an example
of clustering algorithm.

Table 1: Cluster example with customers recommending books.

4
 Book 1 Book 2 Book 3 Book 4 Book 5 Book 6
Customer A X X
Customer B X X X
Customer C X X
Customer D X X
Customer E X X
Customer F X X

As shown in Table 1, any customer shall be classified as a member of a cluster and will
receive recommendations based on preferences on the particular group. As it is shown, book
2 will be highly recommended to Customer F, as well as the book 6 will be recommended to
some extent. Therefore, Customer B, C, and D form a cluster – marked in green – and
Customer A with E will form another cluster towards each other.

2.2 Content-based filtering

Content-based filtering technique (CBF) is a domain-dependent algorithm that relies more
upon the analysis of the attributes of items in order to generate predictions. It is a part of
recommender system that handles with profiles of users that get created in the beginning or
from the past searches depending on the niche. In its process, the engine compares the items
that were already positively rated and recommends to the user accordingly. CBF uses
different sets of models to find similarity between sets of documents in order to generate the
right set of recommendations to the user based on its profile [4]. One of such approach is
with cosine similarity that measures to find contents that matches user profile. It uses vectors
of two items (wc and ~ ws) with attributes that compares in similarity function:
⃗⃗ 𝑐 ∙ 𝑤
𝑤 ⃗⃗ 𝑠 ∑𝐾 ⃗⃗ 𝑖𝑐 𝑤
𝑖=1 𝑤 ⃗⃗ 𝑠
𝑢(𝑐, 𝑠) = cos(𝑤
⃗⃗⃗⃗𝑐 ∙ ⃗⃗⃗⃗
𝑤𝑠 ) = ‖𝑤
⃗⃗ 𝑐 ‖ × ‖𝑤 ⃗⃗ 𝑠 ‖
= .........(1)
√∑𝐾 2 √∑𝐾 𝑤
⃗⃗ 𝑖𝑐
𝑖=1 𝑤
2
𝑖=1 ⃗⃗ 𝑖𝑠

Moreover on vectors, CBF also models the relations between different documents within
a bulk of items such as: Term frequency, naïve Bayes classifier, or neural networks. If the
user profile changes, CBF’s technique still has the potential to adjust itself within a short
period of time. So the profile is often updated automatically when there is a feedback on the
items’ desirability that has been presented to the user. In the figure below is shown a diagram
for CBF and architecture named ‘State of the Art and Trends’ [5].

5
 Figure 2: Content-based architecture diagram[ref]
The advantage of CBF is the lack of data needed from other users in order to recommend
the initial user’s recommendation list. Therefore, depending on what data is needed, CBF
allows transparency that provides explanation for recommended items that lists features of
contents that caused an item to be recommended, especially when it yields a capacity to
recommend new and unknown items, making CBF have a no first-rater problem. An example
of CBF into use is with LIBRA which is a content-based book recommendation system that
uses information about book gathered on the web. Its algorithm uses naïve Bayes classifier
that extracts information from the web and learns from user profiles in there. The system is
able to provide an explanation of each user’s recommendations by listing the highest ratings
of features, giving the users confidence on the recommendation provided to users by the
system itself.
However, automatic assignment of features to items might be insufficient purpose to define
if the analyzed content is not liked by user. Certain drawbacks in CBF may include to certain
users where the user is going to be recommended similar items that has already been rated.
Another drawback may be the overspecialization and obviousness in recommendation to the
user (e.g. ‘Star Trek’ being suggested to a science-fiction fan – accurate but not useful),
making users not wanting algorithms that produce better ratings but rather sensible
recommendations. Thus, the major disadvantage of CBF technique is the need of an in-depth
knowledge and description of the features of the items in profile.

6
2.3 Collaborative Filtering
Collaborative filtering is used mostly when metadata cannot be easily described such as
movies and music, i.e. the idea of collaborative filtering is in finding what the users in a
community share appreciations. Its technique works by building a database of preferences
for items by users – user-item matrix technique (see Figure 3). Next, it matches similarities
of user profiles based on relevant information of interest. To be more exact, the user gets
recommendations to specific items that have not been rated before but were positively rated
by other users with similar interests. Henceforth, RS that are produced by CF can be of either
prediction or recommendation itself. Such potential of similarity between users are named
‘neighbors’. Figure 3 explains further the prediction of items toward the users.

Figure 3: Top N items the user will like the most. Collaborative
filtering[6]

Herlocker et al.[6] came with a new automated collaborative filtering (hereafter: ACF)
systems that predicts a user’s similarity for items or information. So, unlike traditional
content-based information filtering techniques (CBF), filtering is based on human and not
machine, in which it primarily does not depend on the machine analysis for content. It filters
any type of content, be it text, art work, music etc. ACF does not compete with content-based
filtering though. To truly know the collaboration of users, they are more likely to trust a
recommendation when they know the reasons behind that recommendation, so an explanation
is really needed for users to understand the process of ACF, knowing their strength and
weaknesses [6]. So building the explanation facility into recommendation to users can benefit
the user in many ways. It provides a transparency and benefits psychologically speaking,
such as:

7
 • Justification where user is understanding the reason behind the recommendation so it
may decide how much trust it can be put within that recommendation;
• User involvement which allows the user to add knowledge skills to complete decision
processes;
• Education of the user to the processes used in generating a recommendation, so the
user may better understand around the strengths and limitations.
To further understand the concepts of ACF, user first enters the profile of ratings which
processes information and collects a set amount of information from the user, e.g. page-views
and numeric ratings. So an explanation might need to be explained onto what kinds of
preference information were used in a given explanation. In other words, what kinds of data
do the profile consists of? Perhaps the user has not rated enough set of movies to allow the
ACF system to provide accurate recommendations with enough assurance. So what the ACF
does is as in aforementioned explanation of neighborhood, to let the ACF system locate
people with similar profiles. Hence the process that is used to locate other people with similar
profiles is one step closer to success of the collaborative filtering technique. As in the figure
shown above, neighbors selected by the system - usually visually formed in matrices – are
the best predictors for the user’s current information needed; actively searching for new
information will result recommendation to work in the best possible performance; however,
the selected neighbors of the users may not result in the best recommendation for the said
user but they are rather the most similar profiles. The similarity metric is used as a result to
judge the potential neighbors. The final step in this regards, the neighbors’ ratings are
combined to form recommendations. According to Herlocker et al., users can detect instances
in some cases where the prediction is based on a small amount of data, investigating further
to determine if a recommendation is an error or an inactive item.
Although Herlocker et al.[6] study with ACF disbelieves in content-based filtering and
machine learning techniques, they unfortunately failed to provide more mathematical
equations behind the concepts and to truly go in-depth as to how CF works memory-based
and model-based techniques. It is indispensable not to only mention the human-interaction
aspect of collaboration filtering techniques of recommendation, but also onto the machine-
learning techniques with different set of algorithms, as they aid to categorize user within its
neighbors. Consequently, collaboration filtering technique can be divided into two parts:
Memory-based and model-based techniques.

2.3.1 Memory-based technique of Collaborative Filtering

The items that have already been found play a relevant role in neighbor that shares
appreciation with the user. When a neighbor of a user is found, different algorithms will play
an important role in it, in which it has proven to be a success. Therefore, memory-based CF
can be achieved in user-based and item-based techniques.
User-based collaborative filtering technique makes sure to check the similarities between
users that compare their rating on the same item and with that information; it computes the
predicted rating for an item by the active user as an average rating of the item with other
active similar users. Item-based filtering technique bases upon the similarity of items and not
users. It retrieves all items rated by an active user from the user-item matrix, whereas the

8
prediction is made by taking an average of the active users rating the similar item. Pearson
coefficient for correlation further describes the measurement which two variables linearly
relate to one another [7].
∑𝑛
𝑖−1(𝑟𝑎,𝑖 − ̅̅̅)(𝑟
𝑟𝑎 𝑢,𝑖 − ̅̅̅̅)
𝑟𝑢
𝑠(𝑎, 𝑢) = (2)
√∑𝑛 ̅̅̅)
𝑖−1(𝑟𝑎,𝑖 −𝑟
2 𝑛
𝑎 √∑𝑖−1(𝑟𝑢,𝑖 − 𝑟̅𝑢 )
2

In the above equation, s(a,u) denotes that the similarity between two users a and u, ra;i is
rating given to item i by the user a, ra is the mean rating given by user a while n is the total
numbers of items in the user-item.
∑𝑛
𝑖−1(𝑟𝑢,𝑖 − ̅̅̅)
𝑟𝑢 ×𝑠(𝑎,𝑢)
𝑝(𝑎, 𝑖) = 𝑟̅𝑎 + ∑𝑛
(3)
𝑖−1 𝑠(𝑎,𝑢)

In the following formula as seen above, this is called Cosine-based measurement, in which
is mostly used in information retrieval and texts mining to compare text documents. In other
words for the equations above, similarity of items is determined by the similarity of the
ratings of those items by the users who have rated both items, in which such equation further
explains in greater detail as follows:
∑𝑖 𝑟𝑢,𝑖 𝑟𝑣,𝑖
𝑠(𝑢̅ ∙ 𝑣̅ ) = (4)
2 × ∑ 𝑟2
√∑𝑖 𝑟𝑢,𝑖 √ 𝑖 𝑣,𝑖

2.3.2 Model-based technique of Collaborative Filtering

To further state the above-mentioned missing gap by Herlocker et al., apart from using the
ACF with its neighborhood-based techniques, there is another possibility to use the machine
learning technique which is called model-based technique. This technique can be done by the
model building process which as stated, uses machine learning or data mining techniques.
They help tremendously as they can quickly recommend a set of items by using models
similar to neighborhood-based recommender systems techniques. Such examples are
Dimensionality Reduction technique (SVD), Matrix Completion Technique, as well as
regression and clustering.
The use of such algorithms has changed the manners of how recommendation system
works. In greater detail, it went from recommendation by the active users, to recommending
the question of ‘when’ to really consume a product. Examining a few said learning algorithms
used in model-based recommender systems will be further mentioned below.
Clustering: As once mentioned before in the beginning section, clustering algorithm
partitions a set of data into a set of ‘sub-clusters’ in order to discover a meaningful groups
that exists within. Once formed, the opinions of users within a cluster can be averaged out
and makes recommendations for individual users. Therefore, a good clustering will create a
high quality of data between users. A user can have a partial participation in different set of
clusters, making recommendations based on the average across the clusters of participations.
Self-organizing map (SOM) is the most commonly used amongst the different types of
clustering methods. The SOM uses an unsupervised learning method that is based on artificial

9
neuron clustering technique. To further explain, it uses the input nodes which are vectors
with high dimensionality, while the output nodes form a map, such in this example it has 5
by 5 layout in the figure below.

Figure 4: SOM input nodes and output nodes on a 5 by 5 layout[ref].

Regression: It is used when two or more variables are systematically connected by linear
relationship. In other words, it analyzes associative relationships between dependent and its
mirror independent variables. Some examples of the use of regressions contain prediction
and curve fitting which can be useful to identify some sort of trending within dataset, whether
be it linear, parabolic, or other forms.
Matrix completion technique: The essence of matrix completion technique mostly signifies
to predict the unknown values within user-item matrices. K-nearest neighbor is one the
techniques that is most popular amongst collaborative filtering recommendation systems.
They depend upon the historical data of ratings amongst users on items – usually active users;
however, can be a problem due to users not rating most of the items and represented within
matrix, causing the inability to give correct and accurate recommendations.
The model-based and its machine learning techniques, they solve a sparsely problem. The
drawback of the technique is due to its computationally expensive and the memory capacity
usage is extraordinary intensive. The major drawback however, is the cold-state problem
To conclude it all, the collaborative filtering and its techniques have major advantages over
CBF in that it has an ability to perform in domains where there is not much content to begin
with or the content is difficult for a computer system to analyze. One of the advantages is
that it can recommend items relevant to the users even despite not having the content in the
user’s profile. Even despite the successes of CF techniques, their potential problem that
cannot be ignored is the cold-state problem (CS).

10
2.4 Hybrid filtering
Hybrid filtering in recommendation system combines different types of techniques in order
to avoid problems and limitations of the aforementioned pure recommendation systems. The
prototype behind hybrid filtering technique is to combine different set of algorithms in order
to provide more accurate, yet effective recommendations as opposed to only one algorithm
[8]. Using a variety of recommendation techniques will surpass the weakness behind a single
technique by itself in a combined model. Therefore the main reason behind hybrid filtering
is to combine collaborative filtering with content-based filtering in order to improve
accuracy.
To accomplish hybrid filtering, one could categorize in various ways such as:
1. Implement collaborative and content-based methods individually by collecting their
predictions;
2. Integrate characteristics between content-based and collaborative methods;
3. Consolidate a model that integrates both content-based and collaborative
characteristics.
To further state the correlation between CBF and CF techniques within hybrid filtering,
figure 1 shows the methods that CBF and CF recommendations estimate individually, in
order to combine a better recommendation.

CB Recommendation CBF

Figure 5: CBF and CF combined to yield a better recommendation.

Furthermore, figure 2 shows an integration of CBF characteristics into the CF approach.

This especially overcomes the cold start problem in collaborative filtering and problem with
overspecialization of content-based filtering:

CF Recommendation
CF
CBF

Figure 6: CBF characteristics integrated into the CF approach.

In the following figure shows that both CBF and CF can be combined into set of features,
in a unified model in which it improves effectiveness of recommendation process:

11
 Model
CF CBF Recommendation

Figure 7: CBF with CF in a constructed unifying model.

Finally, the last figure of hybrid filtering technique shows the method that integrates CF
characteristics into a CBF approach:

CBF Recommendation
CF

Figure 8: CF features inside CBF approach that yields recommendation.

Content-based filtering system has an allowance towards cold-state items in exchange to
lower accuracy compared to collaborative filtering. Subsequently, collaborative filtering has
a frequent rate of accurate information in exchange to cold-state problem. Hybrid approach
tries to collect these different kinds of information in order for it to get an efficient
recommendation result.
According to Thorat et al, Burke presented taxonomy for hybrid recommendation systems
can be classified into seven following classes [8]:
Weighted: Combines the results of different recommenders to generate a recommendation
list by using a linear formula to integrate scores from each of the techniques. An example is
with P-tango whereas the system consists content-based and collaborative filtering. Initially,
they are given the same weight but do fluctuate when predictions get confirmed or otherwise.
Switching Hybridization: The switching hybrid has the ability to completely avoid specific
to one method e.g. the new user problem of content-based recommender, switching it to a
collaborative recommendation system. The system is sensitive when it comes to strengths
and weaknesses to recommenders. The main disadvantage of switching hybridization is its
featuring of complexities to recommendation systems. DailyLearner uses both content-based
and collaborative hybrid where the content-based is first engaged before collaborative
filtering in situations where content-based system cannot make a recommendation with
sufficient evidence.
Cascade: The cascade hybridization technique applies an improvement in terms of
constructing an order of preferences amongst items. The default technique is refined by

12
another recommendation technique, i.e. the first recommendation technique outputs a list of
recommendations and in turn improved by another recommendation technique.
Mixed: It combines recommendation results at the same time rather than having one
recommendation per item which in return, they will be introduced together. This is based on
merging and presenting multiple rated lists into a single one. An example would be with PTV
system that recommends a TV viewing schedule for the user that combines content-based
together with collaborative systems to form a set of said schedule for the user.
Feature combination: A specified recommendation technique is sent to another
recommendation technique. So the working of the recommender depends on the data that has
an availability to modify by the contributing recommendation technique. Such example is
with Pipper that uses collaborative filter’s rating in content-based system as a feature – e.g.
rating – and it does not rely on collaborative data itself.
Feature augmentation: Similar to feature combination but contributor instead gives more
unique characteristics, making it more flexible. It also requires additional performance and
its functionality from RS. Feature augmentation is superior to its neighbor feature
combination due to number of features it offers to the primary recommender.
Meta-level: When it comes to sparsity problem, meta-level is able to solve it with ease on
collaborative filtering technique by using the model of the first technique as input for the
second one, i.e. the first technique substitutes the data for the second one.

2.5 Machine Learning in Recommender Systems

Recommendation systems in today’s era of time, has been tremendously used within e-
commerce websites, specifically Netflix and Amazon to name a few; however, according to
Cowan et al. [9], the field of RS has its origins created since 1992 with the introduction of a
new RS named Tapestry. After further practices within RS field, researchers studied the use
of algorithms from machine learning (ML), an area of artificial intelligence (AI). Machine
learning has been studied since the late 1950s [9] and as a result, there is a plethora of ML
algorithms (e.g. k-nearest neighbor, clustering, Bayes’ network). As mentioned above, ML
is being used to provide a better use for RS, although ML’s algorithms become more difficult
and confusing for fitting within RS, making it a challenge to tackle the use of ML algorithms
in RSs.

The number or choices and variations for data scientists make it challenging to deal with
ML algorithms. Therefore, one must ask questions and have two main goals:
1. Identify which ML algorithms are most commonly used in RS;
2. To question openly about RS development that might be impacted by data science.

According to the author, 26 publications for ML were retained and ready to be analyzed,
namely from books, conferences, and patents. The results and conclusions from
aforementioned publications are presented in Table 2 for the most used ML algorithms when
it comes to recommendation systems [9].

13
Table 2: Types of Machine Learning algorithms used in recommender systems.
Category Total

Bayesian 7

Decision Tree 5

Matrix factorization- 4
based

Neighbor-based 4

Neural Network 4

As shown in Table 2, Bayesian is the most used ML algorithm when it comes for
recommendation systems, giving 7 out of 26 publications as stated by the author. Both
Bayesian and Decision Trees have similarities when it comes to calculations, as both are a
popular choice.

2.6 The use of Recommender Systems in web applications

Recommendation algorithms are mostly known for their use in e-commerce Web sites,
where they prioritize the customer’s interests to generate a recommendation of items.
Nowadays, plenty of applications prefer to use the items of customer’s purchases and
explicitly rate the customer’s interests, although there can be used a variety of different
attributes, namely items viewed, subject interests, or even their favorite artists.
E-commerce applications and their recommendation algorithms therefore, usually work in
a rather challenging environment, such as:
1. A large retailer (e.g. Amazon, Netflix) often have large amount of datas, millions of
customers with plethora amounts of catalogue items.
2. As a result, many applications require for these large retailers to generate results in
realtime, no more than a split second, all while giving the most quality of
recommendations.
3. New customers stereotypically have low amount of information from the products
they purchased, reviewed, and/or rated (i.e. cold-state situation).
4. Customer’s data is quite volatile to the point where the recommendation algorithms
must respond in an immediate action of generating new information since each
interaction generates customer data.

14
 Table 3 shown for all the known system examples used within RS web applications that
are well-known amongst the RS:

Table 3: System examples within RS and their product goals with models used.

System Product goal Models used Review & facts Languages used
Books and • Collaborative • Pioneer of RS Java,
Amazon other filtering (item- • Virtually sells all kinds of Javascript,
products to-item based) products. Ruby, Python,
• Search-Based • Recommendations are Perl
Methods provided to users on the main
• Cluster Models Web page.
• Purchase/browsing behavior
can be viewed as implicit
rating, as opposed to explicit
rating.
DVDs, • Hybrid • Founded as a mail-order Java, Scala,
Netflix Streaming approaches Dvd, eventually to Javascript.
Videos • Neural streaming.
Networks • User actions watching
• Restricted different items stored.
Boltzmann • Dataset consists of 100
Machines Mio. entries.
• Similarity • Quadruples of <movie,
based on user user, rating, date>
and item
Friends, • Collaborative • A massive social networking PHP, C++,
Facebook Advertisem Filtering website Python, Java,
ents • Matrix • Uses friend Perl
Factorization recommendations as one of
• Main solution: the tools.
Hybrid • Uses structural relationships
Approach rather than ratings data.
Movies • Matrix • Widely known and popular PHP, SQL,
IMDb Factorization online movie platform. ASP.net
• User-based • User ratings are
collaborative complemented with the
filtering Movie Tweetings.

15
3. An in-depth analysis of cold state problem

Oftentimes, the recommender systems are purely based upon the past or present user
ratings for said item, item descriptions using specific keywords within, and user profiles who
share in front of other audience; however, if the information is not readily available for new
users or items, the recommender system will run into a so-called cold-start problem, i.e. cold
state problem. When it approaches in such state, the system does not know what to
recommend, until another user or with enough information, will ‘warm up’, i.e. until it
reaches its threshold of information to start producing recommendations. As an example:
Which product should it be recommended to someone who visits Amazon for the first time?

Figure 5. Cold-state problems (On the left is ‘New Item Problem’. To the
right, ‘New User Problem’)
To further identify, Figure 5 illustrates two sets of cold-state problem: New item problem
is when a new item is added to the specific catalogue and since none has rated the item, it
will never be recommended. Another one being new user problem, whereas a new user has
no ratings, making it hard to predict the said ratings.
As a result, there are several approaches to possibly deal with cold-state problem, namely
utilizing the baselines for cold users, extracting ratings from new users, a system to combine
collaborative filtering with content-based recommenders, or even exploiting the users’ social
networks. With the aforementioned approaches, it is a temporary fix for information to be
gathered and apply to its recommender system, especially in a situation when there are no
‘warm users’ or its period is short-lived; however, in many e-commerce applications, users
or items may remain cold for a long, continuous amount of time or they can get back to a
cold state, leading onto a continuous cold-state (CCS). Such an example can be taken from
Booking.com whereas the user may go for a long break and not return to a site for a while,
leading to a CCS. Furthermore, a warmed, long-term user may become cold, as they change
their needs over the time, wanting to desire entirely different needs over the years. Such users

16
and cases frequently tend to happen in the aforementioned accomodation sites where the user
may only give a visit for business or holiday trips which may be at best, unpredictable for a
system. Therefore, a classical approach to a cold-state problem may fail in case of CCS, since
they assume the user will remain warm after a visit. In this section will be further discussed
about the CCS (continuous cold-state) and a thorough way to give a new user some incentive
reasons to rate through Hossein et al. Ask-to-rate technique which divides in non-adaptive
and adaptive methods.

3.1 User Continuous Cold-State

Kiseleva et al.[10] stated that there are different ways to define the users who become
inactive:
1. Classical CS: This is the case of a new or rare users.
2. Incosistency: User’s interest changes over a period of time and has an incosistent
interests over different times.
3. Identification: User’s identity may act differently, failing to match the same data
from the same user.
The aforementioned cases originate commonly with e-commerce websites. The user may
appear new from a different device, or even another identity logs into the same user, meaning
the system may have a failure of identification matching. Henceforth, some users change
their interests over a period of time, e.g. the movie preferences evolve and changes onto a
different movie niche, user’s travelling has changed in which it tends to happen frequently
or in the case of e-mail marketing, the user becomes inactive and tend to not open, nor click
them. All of such problems arise for collaborative filtering, as well as content-based
approaches, due to user simply not engaging in user ratings, activities and profiles being
empty.

Figure 6: CCS users at Booking.com. Activity levels of two randomly chosen users
over time. (A) Signifies user having an occasional activity throughout a year. (B)
Different personas within a user by making a leisure and a business booking with
no activity in between. Sourced: [10].

17
 Figure 6 (A) shows new users arriving frequently or may appear new when they do not log
in or use a different device altogether. Most users change their interests over time or even
change personas on shorter time. With Figure 6 (B) an example could be simply due to
weather or their travel purpose, wanting to book different type of trips.

3.2 Item Continuous Cold-State

In a uniformal way, CCS problems also defines for items:

• Classical CS: This is the case of a new or rare items’ clicks or impressions by users.
• Incosistency: Item ‘trend’ changes over time; item appeals to different type of users.
• Identification: Failure to match the same data for the item.
New items are in a constant appearance in e-commerce applications. Some of the itemes
are interesting only to niche audience, e.g. books or movies specifically on the topic. Items
may be unpredictable and changes over time, as a new, better version of the same article
takes over the older one, e.g. a new phone releases and takes more sales than the older one.
Item incosistency may derive as a context from news and trendy conversations, defining the
item as a topic shift. When more items are added into the e-commerce site catalogue, or when
even more catalogues are added within, it creates a problem for matching items, hence a
failure of identification. Figure 7 further shows the example of Booking.com with a
continuous cold-state problem within items.

Figure 7: CCS Items at Booking.com. (A) Signifies thousands of properties are

added every month. (B) User ratings over a randomly chosen hotel change over the
year. Sourced: [10].

18
3.3 Non-adaptive methods
The non-adaptive method within ask-to-rate technique makes it possible to present all items
as the same items in new users, regardless of changes within new user’s knowledge.
In 2002, MovieLens researchers came up with new strategies on solving and learning about
new users. The said strategies were focused on the issue of which items to be recommended
to the new users; their outcome through different strategies that were used and experimented
online on their MovieLens datasets, considered user effort and recommendation accuracy
which is related to the user experience. The suggested methods were measured based on
rating prediction accuracy and they were as follows [15]:
Random Strategy: It selects the items randomly which learns about new user preferences
through all the available items. Random strategy is most commonly used for comparisons,
e.g. for items. Unfortunately, the analysis of the rating matrix is not intelligent enough and
shows it through offline and online experiments that it needs more user effort in it.
Popularity strategy: This takes into account on how many users have rated an item. The
items are presented to the new users based on the number of ratings that they have been given
by the users. Furthermore, popularity strategy has an equation: Item 𝑎𝑡 is computed, where
𝑟𝑎𝑡 shows rating.
𝑃𝑜𝑝𝑢𝑙𝑎𝑟𝑖𝑡𝑦(𝑎𝑡 ) = |𝑟𝑎𝑡 | (5)
The point of the equation means to minimize user effort. Ratings however, may be
uninformative since most users like popular items.
To point out more with MovieLens idea of ask-to-rate, in 2008 the researchers extended
its group to improve order of items and extract opinions of new users right away at
registration time. They proposed an offline simulation framework and online experiment with
real users of MovieLens live RS [11, 15].

3.4 Adaptive Method

The items are consistent with each new user’s opinions. So the items get present right away
when the new user attempts to rate an item. It is more controlled more effectively than in
non-adaptive approaches. Adaptive methods take approaches of user’s historical ratings
when it is in its initial state. To deal with the cold-state situation however, there are a few
adaptive approaches such as item-item personalized, information gain through clustered
neighbors (IGCN), naive Bayes, and clustering.

3.5 Methodology used in CS

When it comes to the aforementioned continuous cold-state situation problem, the testing
environment with A/B production was taken advantage for online travel agency e-commerce
sites such as Bookin.com. The A/B testing randomly splits users to see two different versions
of the website – baseline or variant of it. When it comes to baseline, the Kiseleva et al. used

19
a non-contextualized ranker corresponding to the live system [10]. This system is optimized
and therefore is trained to be pressured by a massive volume of traffic. In terms of evaluation
metrics, Kiseleva et al. used in the A/B testing the likes of clicks-per-user and click-through-
rate (CTR). The motivation behind this task was to increase customer engagement, since it
was used with an exploratory task. More clicks and CTR signifies that the user clicks more
on the suggested destinations, interacting more with the system.
Another methodology used in cold state situation when it comes to movie
recommendations can be an example of data used from the ‘MovieLens’ assembled by the
‘GroupLens’ project [11]. Its main purpose is to list movie ratings group by a person.
Approximately, each person would rate at the very least 20 movies. All the MovieLens
observations extract ratings between 1 and 5 respectively. Thus, in order to obtain accurate
actor lists and director information, it heads onto the Internet Movie Database (imdb.com)
by downloading pages and extracting only the top ten actors, eliminating the ones who have
not appeared more than one movie.
The methodology of this MovieLens took 943 people that were made for
recommendations, averaging 85 observations per person in the training set in its entirety, 331
movies into the testing set (out of 1682 of movies in total from imdb at the time) [12]. So
there are three modes identified to test the application: Role of the customer’s purchase of
item or that a customer will both, like and purchase the product (movie), and guess the
customer’s rating on an item that was previously purchased. These tests will be further
elaborated below [12]:
1. Implicit Rating Prediction - This refers to the prediction of data such as the purchase
history itself. Therefore, this purchase cannot indispensably be a total satisfaction by
the customer, but an implicit need for such purchase or desire for an item. In
MovieLens data it predicts that the customer has rated a movie, being comparable to
predicting a customer’s purchase. Implicit rating is appropriate when explicit rating
is not much available but satisfied enough to recommend products that the user is
likely to purchase it.
2. Prediction of Rating itself - The purpose of this is to predict the implicit rating and
items simultaneously. This can be classified into each person or a movie pair that
does not occur in the observation (person as p and m as a movie):
A) 𝑝𝑖 𝑖𝑠 𝑟𝑎𝑡𝑒𝑑 𝑎𝑠 𝑚𝑗 ≥ 4.
B) 𝑝𝑖 𝑑𝑖𝑑 𝑛𝑜𝑡 𝑟𝑎𝑡𝑒 𝑚𝑗 ≥ 4.
Condition of B) could be applied that person i did not rate movie j.
3. Rating Imputation - Rating imputation is a prediction of ratings for items that has
implicit ratings observations. Specifically, a question rises: “Given the person’s seen
movies x, how likely is it to rate it ≥ 4?” The prior knowledge gives in the x which
the person may have watched, meaning that this may be an implicit rating
observation. The rating imputation extracts from the MovieLens data and therefore
predicts the ratings.
In the real-world applications there are data sets which implicit rating observations are
available in large quantities, but the rating items is missing. Henceforth, this rating
imputation fills the gap in the missed values. This is not the first time the rating imputation

20
has been used in recommendation system: “A combination with empirical analysis of
predictive algorithms collaborative filtering” to name amongst many works for
recommendation system.

4. Discussion

To conclude this in a possible CS treatment of data, there are a few algorithms that
individually handle with the case of e-commerce websites. Amazon uses topic diversification
algorithms to improve its recommendations, whereas the system uses collaborative filtering
method to overcome issue with scalability that generates a table of similar items in an offline
mode, by using item-to-item matrix [13]. The system recommends based on the user’s
previous purchase history and throughout the searches the user has made an active seeking
throughout the e-commerce site over the items which helps the system identify the user’s
recent activity. Content-based techniques matches the content to the user’s search
characteristics, as it normally gets based upon the user’s information throughout the search,
ignoring the contributions from other users as with the case of collaborative techniques.
To directly deal with CCS problem, there are traditional approaches when it comes to
dealing with user continuous problem. Recommendations based on social networks is an
interesting new approach as it can supplement missing information previously from the e-
commerce application. As a prime example would be by using Facebook, using
recommendations based on likes are proposed in, even though the identification problem
within a user’s device usage or multiple personnas within a user remains a trouble for now.
When a user clicks throughout the e-commerce website, the browsing behaves and
recommends based on what user is clicking inside the site; however, this still tends to delay
the recommendation system until enough clicks have been made by the user. A more
promising approach to this would be a contextual-based recommendation or as
aforementioned, the content-based recommendations. To begin with a content-based
recommendation therefore, has proven to be effective due to exploitage to find an initial item
based on a single interaction that the user has clicked throughout the site. Context-based
recommendation however, is particularly promising when it comes to solving CCS, by being
based upon the context of the current visitor and the similar behaviors they posses on a similar
context [14]. They define a set of features, e.g. time, location, device to pick a few from
many. These datas cluster users into a context, making the context-based recommendation to
also have created a gap for cold-state of a cold context that the system has never seen the
behavior of that particular user before.
In summary, CF systems yield recommendation based on user-to-user similarity but the
new user encounters a serious problem in the CF approach since it has to acquire some data
about the new user. Hence, the non-adaptive with adaptive approaches were shown for ask-
to-rate technique that exactly deals with new user problem, even though it still has its flaws
with a few methods that have been proposed to deal with CS. In conclusion, there can new
methods in the future that will perhaps deal more with CS problem.
Giving an example, Amazon recently created a new, more incentive ways to engage users.
Such system is called ‘Amazon Prime’ [16] whereas the user subscribes to the Amazon

21
website itself, giving more discounts and incentive reasons for the user to stay in a warm
state.

5. Conclusion

Today, there are plenty of information available on the internet but it is not easy for each
user to find relevant information in short amount of time. In order to overcome this problem,
the recommendation system was introduced. So in this paper is presented a RS that open new
opportunities to retrieve personalized information on the internet. The problems and solutions
are presented in the most known recommendation system models and techniques that are
analyzed in greater details. Various algorithms and machine learning are mentioned for their
quality and performance with a few examples, all convenient for recommendation system.
This paper also discussed two traditional (i.e. content-based and collaborative filtering)
techniques for recommendation which highlighted their overall strength and performances,
together with their challenges and ways to overcome.
However, the big issues unfortunately is in fact that user interests and taste changes with
time – as a result, social networks are the first and most active place to be notified with such
changes. Furthermore, these recommendation system techniques are limited in the way they
operate – e.g. an e-commerce site can yield recommendations only within purchases made or
products browsed by users. So the developers have a good chunk of data from their own
monitoring their incentive e-commerce platform (e.g. Amazon prime) and social networks
which increases user’s activity, creating precisive and up-to-date user profiles that can be
used for recommendation such as ads, consumer goods, and travel to name amongst many
use cases. Another problem within recommendation system is data sparsity whereas an e-
commerce site has a big catalogue but very few purchases and this especially is problematic
to a new user which would be virtually impossible to recommend anything, hence why
mentioning the context-aware recommendation system may help deal with the situation in a
cold state. So when a user visits the website, its browsing behavior is used to estimate its
intent of after a few clicks through the website – be it searching for different products or user
profiles. Another good solution to the cold state situation is the use of data gathered through
social since the user profile is – most of the time – already detailed.

References
[1]. Mayer-Schönberger, V., & Cukier, K. (2013). Big data: A revolution that will
transform how we live, work, and think. Houghton Mifflin Harcourt.

[2]. Arora, S. (2016). Recommendation engines: How Amazon and Netflix are winning
the personalization battle. MarTech Advisor.

[3]. Isinkaye, F. O., Folajimi, Y. O., & Ojokoh, B. A. (2015). Recommendation systems:
Principles, methods and evaluation. Egyptian Informatics Journal, 16(3), 261-273.

22
 [4]. Pazzani, M. J., & Billsus, D. (2007). Content-based recommendation systems.
In The adaptive web (pp. 325-341). Springer, Berlin, Heidelberg.

[5]. Di Noia, T., & Ostuni, V. C. (2015, July). Recommender systems and linked open
data. In Reasoning Web International Summer School (pp. 88-113). Springer,
Cham.

[6]. Lops, P., De Gemmis, M., & Semeraro, G. (2011). Content-based recommender
systems: State of the art and trends. In Recommender systems handbook (pp. 73-
105). Springer, Boston, MA.

[7]. Herlocker, J. L., Konstan, J. A., & Riedl, J. (2000, December). Explaining
collaborative filtering recommendations. In Proceedings of the 2000 ACM
conference on Computer supported cooperative work (pp. 241-250).

[8]. Melville, P., Mooney, R. J., & Nagarajan, R. (2002). Content-boosted collaborative
filtering for improved recommendations. Aaai/iaai, 23, 187-192.

[9]. Thorat, P. B., Goudar, R. M., & Barve, S. (2015). Survey on collaborative filtering,
content-based filtering and hybrid recommendation system. International Journal of
Computer Applications, 110(4), 31-36.

[10]. Portugal, I., Alencar, P., & Cowan, D. (2018). The use of machine learning
algorithms in recommender systems: A systematic review. Expert Systems with
Applications, 97, 205-227.

[11]. Kiseleva, J., Tuzhilin, A., Kamps, J., Mueller, M. J., Bernardi, L., Davis, C., ... &
Hiemstra, D. (2016). Beyond movie recommendations: Solving the continuous cold
start problem in e-commercerecommendations. arXiv preprint arXiv:1607.07904.

[12]. Ekstrand, M. D., Kluver, D., Harper, F. M., & Konstan, J. A. (2015, September).
Letting users choose recommender algorithms: An experimental study.
In Proceedings of the 9th ACM Conference on Recommender Systems (pp. 11-18).

[13]. Schein, A. I., Popescul, A., Ungar, L. H., & Pennock, D. M. (2002, August).
Methods and metrics for cold-start recommendations. In Proceedings of the 25th
annual international ACM SIGIR conference on Research and development in
information retrieval (pp. 253-260).

[14]. Ziegler, C. N., McNee, S. M., Konstan, J. A., & Lausen, G. (2005, May). Improving
recommendation lists through topic diversification. In Proceedings of the 14th
international conference on World Wide Web (pp. 22-32).

[15]. Adomavicius, G., & Tuzhilin, A. (2011). Context-aware recommender systems.

In Recommender systems handbook (pp. 217-253). Springer, Boston, MA.

23
 [16]. Nadimi-Shahraki, M. H., & Bahadorpour, M. (2014). Cold-start problem in
collaborative recommender systems: Efficient methods based on ask-to-rate
technique. Journal of computing and information technology, 22(2), 105-113.

[17]. Dunn, Jeff. “Amazon Keeps Giving Goodies Away to Prime Members Because It
Pays off in the End.” Business Insider.

[18]. Elahi, F. B. M. M. (2019). Cold Start Solutions For Recommendation Systems. IET.

24

View publication stats