International Journal of Computer Science and Information Security (IJCSIS),

Vol. 15, No. 9, September 2017

AN EFFICIENT TRUST EVALUATION USING FACT-FINDER

TECHNIQUE
K.T.Senthil Kumar1 and Dr.R.Ponnusamy2
1
Part-Time Research Scholar in Computer Science, R&D Department, Bharathiar University, Coimbatore, India.
2
Professor, Department of CSE, Sri Lakshmi Ammal Engineering College, Chennai, India
ABSTRACT I. INTRODUCTION
Provide individualized suggestions The eventual characteristic of
of data or products related to users’ needs internet 2.0 is the free generated content by
by Recommender systems (RSs). Even users. Users have the power to reveal their
if RSs have created substantial progresses Opinions in Online Social Networks OSNs
in theory and formula development and such as Facebook, Wikis and Twitter. In
have achieved many business successes, a fact, an exponential growth of information
way to operate the wide accessible info in has become available to users.
online social Networks (OSNs) has been Accordingly, two cases are imposed: one
mainly overlooked. Noticing such a gap in is the difficulty for users to find contents
the existing research in RSs and taking that are relevant to their own interest
into account a user’s choice being greatly among vast amount of alternatives. The
influenced by his/her trustworthy friends other is the demand to a modern technique
and their opinions; this paper proposes a, which can provide personalised
Fact Finder technique that improves the recommendations by exploiting
prevailing recommendation approaches by information in the current environment of
exploring a new source of data from World Wide Web.
friends’ short posts in microbloggings as In fact, traditional Recommender
micro-reviews.Degree of friends’ Systems (RSs) play a very important role
sentiment and level being sure to a user’s in providing recommendations and those
choice are known by victimisation they are deployed within
machine learning strategies as well as the business, like Amazon, Netflex and
Naive Bayes, Logistic Regression and EBay Typically, there are two main
Decision Trees. As the verification of the techniques of RSs: Collaborative Filtering
proposed Fact finder, experiments recommenders (CF) and Content-Based
victimisation real social data from Twitter recommenders (CB) [1], [2]. In CF
microblogger area unit given and results techniques, recommenders can predict
show the effectiveness and promising of relevant items for an active user by
the planned approach. utilising his previous history of ratings
from similar users as neighbours [3] or
Index Terms — Recommender systems, similar things [4]. On the opposite hand,
machine learning, trust, sentiment analysis, CB recommenders enrich recommendation
microblogging, Fact-finder by building user-item rofiles supported the
foremost vital options of item contents [5],
[6]. Usually, this technique depends on
things containing matter data, for instance,
content of websites recently, within
the try of rising ancient RSs; some studies

10 https://sites.google.com/site/ijcsis/
ISSN 1947-5500
 International Journal of Computer Science and Information Security (IJCSIS),
Vol. 15, No. 9, September 2017

enhance the recommendation by including rating history. We propose Fact Finder

trust relations. Trust increases the ability technique. It is based on the assumption
of RSs to approach more trusted users and that users tend to be influenced by their
as a consequence more reliable products to friends’ opinions even if they have
be suggested [7]–[9]. Another different interests. We have a tendency to
methodology proposed to empower argue that OSNs, microbloggings in
recommendation by extracting the particular, will be an upscale supply of
sentiment data from long reviews that data to change recommendation. Our
users have written concerning merchandise analysis interest is, for an energetic user, to
[10]–[12]. harness his/her friends’ sentiments about
However, there are several products by exploitation the trust degree
limitations which still appear in the between the user and his/her friends in
aforementioned approaches. First of all, social network. We explore whether, for a
the user-item rating matrix in CF user, his/her trusted friends’ posts could
techniques suffers from the well-known also be thought of as short reviews to
problem of sparsity as people tend to rate empower recommendation.
few items, and this produces less dense of To the Simplest of our information, this
the available ratings in RSs as pointed out can be the primary work using sentiment
in [4]. Further, the challenge of the cold- and trust from microbloggings to generate
start problem when new users or items do personalized recommendations. The
not have any rating history. Traditional CF contributions of this paper are:
and CB recommenders assume the Employing OSNs as a replacement supply
existence of sufficient amount of ratings or of knowledge so as to use the short posts
content information in order to generate messages as micro-reviews in the
powerful recommendation but this is not projected Fact-Finder recommendation
true in many cases [1]. Second, most of framework.
trust-enhanced RSs are not practical as 1) Inferring multiple score ratings from
they do not replicate the influence of real friends’ posts in microbloggings by
social connections in providing victimization sentiment analysis technique,
recommendations. In globe, we have a as these posts are unit short and embrace
tendency to tend to trust our friends’ informal use of language.
opinions about books, movies and 2) Using intercommunication between
restaurants. Third, recommendation friends as the trust indicator to the
extracted from users’ product reviews importance of friends’ opinion to a user.
don’t use any benefits which might be 3) Improving the prediction performance
controlled to change recommendations using different machine learning
from friends in OSNs such as Twitter, classification algorithms, in particular, for
since vast amount of information and new users.
opinions are available in such networks. The rest of this paper is organised as
In this paper, we aim to propose a follows. In Section II, we provide an
solution to the above problems and model overview of some major studies and
a recommender which can involve a user’s approaches for recommender systems. In
OSNs to draw the user’s preferences even Section III, we explain the environment of
in the case that he/she does not have any social network data that we utilise. Section

11 https://sites.google.com/site/ijcsis/
ISSN 1947-5500
 International Journal of Computer Science and Information Security (IJCSIS),
Vol. 15, No. 9, September 2017

IV details the problem we study in this involved social connections data in

paper. Our approaches in predicting ratings providing recommendation by assigning
are presented in Section V. The results of social regularization terms in order to
the experimental analysis are presented in constraint matrix factorization objective
Section VI, followed by the conclusion function. They assumed that friends rate
and future work in Section VII. products and hence they used PCC and
II RELATED WORK VCC to measure similarity as intermediate
step.
In this section we tend to highlight Second, additional studies
related important approaches: 1) ancient have targeted on trust-enhanced
collaborative recommender, 2) trust- recommenders. Some studies applied trust
increased recommender, and 3) reviews- by building trust net- work supported the
based recommender. belief that users will get additional
First, the standard collaborative accurate recommendation from folks they
filtering approaches are either memory- trust [18], [19]. These styles
based or model-based. These strategies are of strategies used direct evaluations of
based on the rating history from users trust from users. Golbeck et al. [18]
within the memory-based strategies, propagated trust from trust network so-
similarity computation is a primary part. referred to as net Of Trust
They use a heuristic utility of similarity WOT. Solely friends whose
between users’ vectors like Pearson trust analysis exceeds a threshold are going
Correlation Coefficient (PCC) or to be concerned in recommender encounter
trigonometric function similarity measure Recommendations square
(VCC) [1], [5], [13], [14]. On the opposite measure obtained by weighted average of
hand, the model-based strategies used ratings alongside the
machine learning models to predict trust price victimisation
product ratings [15], [16]. For example, Film Trust dataset. In another context,
Sarwar et al. [4], [17] implemented Massa et al. [19] used trust to filter the set
clustering algorithms to identify groups of of neighbours and solely their things
customers who rated similar products and would be thought of in predicting ratings
these clusters can be seen as likeminded to an vigorous user. once filtering
neighbours. Since k clusters are created, neighbours, they applied the standard
recommendation prediction can be recommendation algorithm. The
computed by averaging the ratings in that experiments were supported Opinion
cluster. Miyahara and Pazzani [5] dataset that contains each user’s ratings
proposed a RS based on Naive Bayes and also the trust values from users
classifier and they only considered items towards one another.
which co-rated between users. They
manipulated two classes: like and don’t Third, recent researches are done to
like and features are selected in a pre- exploit the sentiment in the textual reviews
processing step. Recent proposals focused to augment ratings in collaborative
in the accuracy of predictions such as recommenders [11], [12], [20]. Authors in
matrix factorization for collaborative [11] tried to enhance the RSs
filtering. The approach proposed in [8] by leverage topic and sentiment data at

12 https://sites.google.com/site/ijcsis/
ISSN 1947-5500
 International Journal of Computer Science and Information Security (IJCSIS),
Vol. 15, No. 9, September 2017

sentences level. They inferred ratings from network as it is in the user-item rating
textreviews written by matrix. On the other hand, review based
users regarding restaurants in multi-point recommenders require a user to write
rating scale instead of solely positive or reviews and rate products to generate the
negative polarities. They applied text suggestions. Unlike existing studies, our
regression model to estimate scaled novel approach fact-finder overcomes the
sentiment ratings. They’re the need of ratings or written reviews by users
primary WHO integrated and reflects the real hidden social trust
the helpful data in reviews into RSs. relations. In our work we personalise
Lenug et al. [12] planned a probabilistic recommendations from microbloggings
sentiment logical thinking framework. using sentiment analysis and trust between
They applied friends.
Natural language techniques to compute III MICROBLOGGING SERVICES
sentiment orientation in reviews. They
designed their rating inference model In this section we have a tendency to
supported on the Naive mathematician introduce our target social network
classifier. Then, they integrated between Twitter. Users with in the microblogger
the logical thinking ratings from reviews Twitter will publish short posts in 140
and a CF algorithmic program to character limit questionable tweets. Today,
extend users’ preferences and achieved Twitter users will generate over 300
encouraging results. Esparza et al. in [21] Million tweets each day [22] about
investigated how to obtain different topic and interest. For example,
recommendation from online people can generate brief posts about their
microblogging services. They proposed a personal experience in reading books,
solution to exploit short posts written by watching movies, breaking news or even
users as product reviews. These posts are the release of new electronic gadgets.
used to build user-item profile. Then a Additionally, users have the selection to
query search algorithm is applied to determine relationships among every other
retrieve relevant item profiles based on a for social links, seeking info or
twitter-like review service called distinguishing following/followers friends.
blipper.com. This study is similar to our
work in using microblogging as a source
of recommendation.

Some inherent drawbacks still have not

been solved in the above mentioned
methods. Most of these approaches require
users to produce some structured data first
such as trust evaluations and ratings to
allow the corresponding systems to work
properly. In fact, this is not practical and
usually not available. Nevertheless, the
Fig.1Intercommunication with User’s
weaknesses of sparsity and cold start
and friends
problems appear in the case of trust

13 https://sites.google.com/site/ijcsis/
ISSN 1947-5500
 International Journal of Computer Science and Information Security (IJCSIS),
Vol. 15, No. 9, September 2017

drawing the user’s taste and preferences

isn’t offered and most of the existing RSs
Measuring the various levels of hidden and algorithms cannot give the individualized
subjective trust relationships between suggestions.
friends in Twitter is crucial in our analysis. What usually folks do once they measure
Therefore, we developed a tool to lack of information is to raise their friends
mechanically collect social network as they trust them. People also tend to
information by using Twitter API. This point out interest and curiosity in things
tool extracts the specified interactions (movies, books, restaurants ...etc) that their
between friends, and it is known as Twitter friends like. Being influenced by friends’
Interaction Extractor (TIE), additional style could be a common feature within the
details in Section VI. It is important to real world. For the example in Fig. 1, the
high light that we access only accounts active user has five friends who
that available to public and not any broadcasted different messages, for
protected accounts for private security. example, Let us assume the domain of
however, some regulations that Twitter movie recommenders, then the
service applies increase the challenge of investigation of his friends’ circles shows
getting such social network information, that Friend1, Friend2 and Friend3 have
for instance, the rate limit of accessing and some knowledge about movies. The
requesting information from Twitter challenge is that the active user definitely
service and the dynamic change in the has different relationships with his peers.
relations and contents. Moreover, some We need to analyse these social ties and
users choose to apply more privacy understand that friend that the active user
constraints on their accounts in order to could trust his opinion the foremost. To do
avoid their info to be revealed to general so, it’s vital to require in to consideration
public. the communication behaviours between
the active user and his friends as a trust
indicator. Samples of these interactions are
IV METHODOLOGY the action of resending messages from his
friends (RE) and saving his friends’ posts
Fact-Finder Cluster Technique in favourite list (FV). It is also highly
We believe that the social relations needed to search out the extent of
showing among users and friends in OSNs sentiment orientation in friends’ posts
influence users’ buying behaviour. As the about regarding movies and contemplate
quantity of data and opinions regarding the informal use of languages and icons.
products and services increase and diverse Next section explains how to compute the
additional and additional, harnessing implicit social trust, and then sentimental
friends’ opinions and incorporating trust analysis of the Post (or) comments is
relations between them to enhance introduced in Section IV-B.
recommendation are getting a vital need.
Fig. 1 is an example for an online active
user who has not experienced lots of items
or a completely new user in a retailer
website. During this state of affairs,

14 https://sites.google.com/site/ijcsis/
ISSN 1947-5500
 International Journal of Computer Science and Information Security (IJCSIS),
Vol. 15, No. 9, September 2017

A. Implicit Social Trust

In this section, we will observe some

necessary communication activities
between peers to indicate how people trust
one another in OSNs. We tend to believe
that the interactions between friends in
OSNs will indicate what proportion trust
they’ll hold towards completely different
friends. Some actions show however
friends perceive each other such as re-
tweeting, mentioning others, favouring
others’ posts and number of followers
[23]. Communication activities in Twitter
that we consider in this paper are defined
as: the action of re-tweeting which implies
that a user re-sends a tweet to all his/her
friends to show the interest, and will be
denoted as RT. We would like to compute
the trust relation between user u and one Fig2. Proposed System
friend f among the cluster of friends F
since f ∈ F.
Due to the fact that people interactions
Intuitionally, trust is identified as a vary over the time and relations are not
normalised average static, we define the periods of times as

, T = {t1, t2, ...tW } and then the same

Trustu, f =
computation of trust in equation is applied
for each time period tj ∈ T ,
where we denote trust between user u and
friend f as u, f trust . And u, f RT is the ,
Trustu, f (t j) =
number of re-tweeted messages done by u
to friend f in a given period of time which Based on the above equation, we can
is the total re-tweeted messages done by u detect the trust between u and f over all
to all friends in Cluster F denoted as u, f periods of time T as,
RT in that period of time.
TRUSTu, f (T) = ∑ (trust) ( tj)

B. Sentimental Analysis using Natural

language processing

How will we have a tendency to

learn the probabilities P(c) and P (fi |c)?
Let’s first take into account the utmost
change estimate. We’ll simply use the

15 https://sites.google.com/site/ijcsis/
ISSN 1947-5500
 International Journal of Computer Science and Information Security (IJCSIS),
Vol. 15, No. 9, September 2017

frequencies within the knowledge. For the systems based on five or ten score rating
document previous P(c) we ask what scale. This can be further explained by,
percentage of the documents in our Positive and negative words are tokenism
training set are in each class c. Let Nc be and future improving their levels.
the number of documents in our training
data with class c and Ndoc be the total V. ITEMS INFERENCE RATINGS
number of documents. The crucial step in RSs is to generate
(en.wikipedia.org/wiki/Emoticon) such as prediction of ratings. We develop two
the emoticon” :)” and different techniques one based on heuristic
abbreviation”OMG”. Positive and negative foundation to predict ratings in Section V
intensifiers are also counted such as (A) and the second is to apply machine
“cooool”. Binary vector of options is learning models in Section V (B).
enforced manually for each micro-review A. Heuristic Prediction
and any existence of the sw will be given The group of friends who have
worth one even positive or negative, opinion of the items then among the all
otherwise it will take the zero value. friends
1) It is important to say that trust between
In contrast to the binary sentiment ways one user u and his
that offer solely the polarity of reviews as friend f may be not bidirectional, hence:
negative or positive, we want to get if u → f and f → u then trust u→f ≠ trust
additional precise sentiment analysis that f→u
describes over easy like or dislike ways.. 2) If (trustu,f1 ) > (trustu,f2 ) then, srf1
For a given set of micro-reviews every is contributes more to Ru, i than srf2 .
represented by a set of sentiment words 3) MI Nf ∈ TF (srf,i ) ≤ Ru,i ≤ M AXf ∈
We want to infer sentiment rating sˆr to TF (srf,i )
hold a class of ratings for example, since
our goal is to allocate a rating to describe B. Models-Based Prediction
the strength of an opinion in micro- Used to validate all of the information’s
reviews. Now we will infer sentiment of all users and friends can contribute to
ratings by aggregating all the existence of products ratings estimation by using
positive sw’s normalised by the full classification learning algorithms.
variety of existence mentioned options Naive Bayes classification, Logistic
previously either positive or negative, regression and Decision tree. The
similar methodology is used in [11] but importance of these algorithms comes
they worked solely at sentence level while from yielding good result in different
not together with the special language domain and the availability of the related
option. The subsequent equation illustrates software tools [24]. Applying these
however we have a tendency machine learning methods on our social
to cipher the reasoning sentiment rating sr data SD = {trust u, f , srf,i }
from mr: needs to identify the relations between
(, ) features terms by IR which indicates the
Sr f,I = (, ) (, ) friend’s opinion (sr)on user-rating about
Number of the class categories used in
recommendation, for example, some IR( u , f , i ) = g (trustu , f , srf , i )

16 https://sites.google.com/site/ijcsis/
ISSN 1947-5500
 International Journal of Computer Science and Information Security (IJCSIS),
Vol. 15, No. 9, September 2017

Where g is an unknown function we need I ( p ) = p i log2 pi

to define. And then we aggregate the
impact function I R to obtain the final user VI. EXPERIMENTAL ANALYSIS
ratings as: A. Dataset
The three well-known machine learning The dataset that we like should contain
models are used to represent or friends revealed posts (micro-reviews) from
approximate the unknown function g as the OSNs and trust relations besides the classic
output will be one of the class ratings and users and items ratings info. In fact, there is
feature vector we examined the no adequate dataset with the predefined
classification models on the social data requirements since the available datasets
dataset SD to predict one of five nominal contain only ratings, only reviews or ratings
classes of ratings Extreme Like, Like, with reviews. Given the absence of friends
Neutral, Dislike, Extreme Dislike. In fact, information, therefore, we tend to designed
the core engine behind these three a package tool TIE to scrawl Twitter and
classification algorithms is quite different. prepare the friends information using the
Next paragraphs describe brief details Twitter API for JAVA (contact the authors
about each. for the collected dataset).
We could collect social data about 111
P c X P c X for i, j ∈ {1, .. m}
users’ information as follows: firstly, we
There are some premises to apply NB. It is
randomly choose movies form the popular
simple technique to use and tends to be
movie lens dataset after that we used the
optimal for particular domain classes with
search tool TIE to gather information about
highly independent and irrelevant features.
a person who posted a tweet about the
Moreover, the probabilistic nature of NB
chosen movie such as name and twitter ID.
allows it to handle missing values [25].
Secondly, since we were able to allocate the
We describe two more algorithms might
publisher information then we start
achieve better results.
to detect the re-tweeting messages activities
A decision tree model is also applied.
It is a nested set of rules that used to split rate between this person and his friends. On
the data. This recursive algorithm the other hand, we used three
constructs a tree structure automatically B. Metrics
starting from root features and ending with
leaf nodes. When splitting the data a To compare the performance between
decision rule is applied for every feature the three applied machine learning
then the feature that minimises the cost
function is chosen to build tree branches. classification algorithms we applied
The leaf node at the end of each branch is different metrics. We used Accuracy metric
a class. There are many decision tree to indicate the percentage of the correctly
algorithms in the literature, in this work classified instances in the test set. However,
we adopt C4.5 algorithms. The metric that this metric is not enough because it is not
used to measure the best splitting of data sensitive to class distribution or the chance
by C4.5 algorithm is called the information
of being correct. Hence, we also used the
gain IG measurement derived from the
dataset itself to split the tree branches. Let standards evaluation measurements that are
pi is the probability that a subset of SD widely used in information retrieval and
labeled by ci , then: classification such as Precision and Recall
defined in equations (17), (18)

17 https://sites.google.com/site/ijcsis/
ISSN 1947-5500
 International Journal of Computer Science and Information Security (IJCSIS),
Vol. 15, No. 9, September 2017

respectively. These two metrics test the 8

accuracy of classification algorithm in 7
6

Accuracy
predicting ratings [25]. 5
4 Precision
3
2
Precision = 1 Recall
0
F-measure

Recall =

F-measure is considered as the harmonic Rating Classes

mean between the two metrics Precision and
recall to overcome any conflict between
Fig. 3. Precision, recall and F-measure
them. It is given as follow:
for each ratings class produced by
decision trees.
.
F- Measure = 2.
c. Evaluation Result
Experiment is applied to evaluate the
For testing, we randomly split the dataset
recommendation classification algorithms
into ten non-overlapped folds to apply 10-
accuracy given the computed trust and
fold cross-validation. The experiments are
sentiment values derived from the
repeated on the ten folds. Every fold is
collected data described in Section VI-A.
used as test set and the rest nine folds used
F-measure of NB, logistic regression and
as training set. Our results are computed
decision tree were 0.57, 0.67 and 0.72
based on the average of all the ten folds
respectively. This means that decision tree
Models.
has power of prediction higher than the
We also applied the statistical accuracy
two others algorithms. It is clear that NB
metrics such as Mean Absolute Error
classifier has the worst accuracy
(MAE) to evaluate the recommendation
percentage while the best performance is
algorithm. It is the most widely used and
given by decision tree it can correctly
acceptable in the recommendation
predict 72% of test ratings. More
community because it is easy to apply and
specifically, Fig. 3 sheds the light on these
we can interpret comparisons directly.
metrics results using decision trees
MAE is defined,
according to each class category. There is
where N is the size of the test set, and ru,i
more determination about negative classes
is the rating assigned by user u to movie i,
Dislike and ExtermeDislike than positive
and rˆu,i indicates the rating estimated by
classes. Meanwhile the Neutral class gains
the proposed recommendation algorithm.
the lowest accuracy and this may due to
Obtaining small results of MAE shows
the ambiguous nature with this class since
more accurate performance of the system.
it holds uncertainty about opinion. It is
difficult to know whether Neutral class is
Neutral or Extreme Like which are
closer to which polarities (negative or
semantically relevant rather than going
positive) unless further contextual
further distance classes such as Dislike or
information is included such as
ExtermeDislike.

18 https://sites.google.com/site/ijcsis/
ISSN 1947-5500
 International Journal of Computer Science and Information Security (IJCSIS),
Vol. 15, No. 9, September 2017

demographic information or more trust

indicators are used.
Algorith Accuracy Precision Recall F-measure Algorithms MAE
ms B1 1.292
B2 1.735
Naïve 57.65% 0.58 0.58 0.57
Bayes B3 1.278
Logistic- 67.6% 0.66 0.68 0.67 B4 3.236
Regressi ISTS₁ 1.20
on
ISTS₂ 0.836
Fact- 72.1% 0.73 0.72 0.72
finder

VII CONCLUSION
d)Recommendation Results In this paper, we have tendency to
explore the potential of social info derived
We will evaluate the performance of our
from microbloggings as a supply of user
approach having group of friends. In this
experiment, we assume the existence of relevant recommendations. In distinction
totally 12 new users. Each user is assigned to ancient RSs which are based mainly on
a group of friends of size 10 sampled from structured data, we investigate the data
the dataset SD. Moreover, we applied a comes from the current web environment.
threshold of the lowest acceptable trust We propose the approach Fact finder that
value. Only friends who exceeding this can exploit two factors from OSNs: the
threshold will be involved in
sentiment orientation in friends posts
recommendation process. We chose the
value of the threshold to be 0.17 as this regarding sure things and therefore the
value of trust is the information gain in trust relations between friends. Our
decision tree model. This threshold will be evaluations are applied on real social
applied for all baselines. After selecting knowledge from Twitter. The results show
the trusted friends, this local community that the these short and inconsistent posts
will be treated as user’s neighbourhood. can empower the users preferences data in
Table III reports different values based on
the average of MAE from friends groups particular when no preferences of history
of the 12 users. B4 achieved the highest were available. Several machine learning
error among all the methods. This is classification algorithms were accustomed
because it is based on the traditional classify a score rating, and the tree
collaborative filtering method which decision model performs the best accuracy
requires both common ratings between metrics results. In the future challenge, we
friends or items, and user average ratings.
have tendency to believe that user’s taste is
Thus, traditional collaborative filtering
lead to poor recommendation in new user absolutely very important in
situations. Due to the large error obtained personalisation. Therefore we have a
by B4 we excluded it in the next tendency to attempt to fuse user’s own
experiment. It can be observed that both of preferences - once exists - with social info
our methods I ST S1 and I ST S2 gained comes from OSNs to counterpoint and
the smallest error results 1.2 and 0.836 augment collaborative filtering
respectively.
recommenders.

19 https://sites.google.com/site/ijcsis/
ISSN 1947-5500
 International Journal of Computer Science and Information Security (IJCSIS),
Vol. 15, No. 9, September 2017

REFERENCES ACM International Conference on Web

[1] G. Adomavicius and A. Tuzhilin, Search and Data Mining, ACM,
“Toward the next generation of 2011, pp. 287–296.
Recommender systems: A survey of the [9] Y. Kim and R. Phalak, “A trust
state-of-the-art and possible extensions,” prediction framework in rating-based
IEEE Transactions on Knowledge and experience sharing social networks without
Data Engineering, vol. 17, no. 6, pp. 734– a web of trust,” Information Sciences,
749, 2005. 2012.
[2] F. Ricci, L. Rokach, and B. Shapira, [10] B. Pang, L. Lee, and S.
“Introduction to recommender systems Vaithyanathan, “Thumbs up?: Sentiment
handbook,” Recommender Systems classification using machine learning
Handbook, pp. 1–35, 2011. techniques,” in Proc. the
[3] P. Resnick, N. Iacovou, M. Suchak, P. ACL-02 Conference on Empirical Methods
Bergstrom, and J. Riedl, “Grouplens: An in Natural Language
open architecture for collaborative filtering Processing, vol. 10, 2002, pp. 79–86.
of netnews,” in Proc. the 1994 ACM [11] G. Ganu, N. Elhadad, and A. Marian,
Conference on Computer Supported “Beyond the stars: Improving
Cooperative Work, ACM, 1994, pp. 175– rating predictions using review text
186. content,” presented at 12th
[4] B. Sarwar et al., “Item-based International Workshop on the Web and
collaborative filtering recommendation Databases, 2009.
algorithms,” in Proc. the 10th [12] C. W.-K. Leung, S. C.-F. Chan, F.-L.
International Conference on World Chung, and G. Ngai, “A
Wide Web, ACM, 2001, pp. 285–295. probabilistic rating inference framework
[5] M. Pazzani and D. Billsus, “Learning for mining user preferences
and revising user profiles: The from reviews,” World Wide Web, vol. 14,
identification of interesting web sites,” no. 2, pp. 187–215, 2011
Machine Learning, vol. 27, no. 3, pp. 313– [13] J. A. Konstan, B. N. Miller, D. Maltz,
331, 1997. J. L. Herlocker, L. R. Gordon,
[6] L. Si and R. Jin, “Flexible mixture and J. Riedl, “Grouplens: applying
model for collaborative filtering,” in collaborative filtering to usenet
Proc. Machine Learning-International news,” Communications of the ACM, vol.
Workshop then Conference, vol. 20, no. 2, 40, no. 3, pp. 77–87, 1997.
2003, p. 704. [14] X. Su and T. M. Khoshgoftaar, “A
[7] H. Ma, I. King, and M. R. Lyu, survey of collaborative filtering
“Learning to recommend with social trust techniques,” Advances in Artificial
ensemble,” in Proc. the 32nd International Intelligence, p. 4, 2009.
ACM SIGIR Conference on Research and [15] K. Miyahara and M. J. Pazzani,
Development in Information Retrieval, “Improvement of collaborative
ACM, 2009, pp. 203–210. filtering with the simple bayesian
[8] H. Ma, D. Zhou, C. Liu, M. R. Lyu, classifier,” IPSJ Journal, vol. 43, no.
and I. King, “Recommender systems with 11, pp. 3429–3437, 2002.
social regularization,” in Proc. the Fourth [16] X. Su and T. M. Khoshgoftaar,
“Collaborative filtering for multi-class

20 https://sites.google.com/site/ijcsis/
ISSN 1947-5500
 International Journal of Computer Science and Information Security (IJCSIS),
Vol. 15, No. 9, September 2017

data using belief nets algorithms,” in Proc. methods for recommender systems,”
International Conference on Tools with Recommender Systems Handbook,
Artificial Intelligence, IEEE, 2006, pp. Springer, 2011, pp. 39–71.
497–504. [26] S. L. Cessie and J. V. Houwelingen,
[17] B. M. Sarwar, G. Karypis, J. Konstan, “Ridge estimators in logistic regression,”
and J. Riedl, “Recommender systems for Applied Statistics, pp. 191–201, 1992.
large-scale e-commerce: Scalable [27] L. Rokach and O. Maimon, “Top-
neighborhood formation using clustering,” down induction of decision trees
in Proc. the Fifth International Conference classifiers-a survey,” IEEE Transactions
on Computer and Information Technology, on Systems, Man, and Cybernetics, Part C:
vol. 1, 2002. Applications and Reviews, vol. 35, no. 4,
[18] J. A. Golbeck, “Computing and pp. 476–487, 2005.
applying trust in web-based social [28] P. Victor, M. D. Cock, and C.
networks,” 2005. Cornelis, “Trust and recommendations,”
[19] P. Massa and P. Avesani, “Trust- Recommender Systems Handbook,
aware recommender systems,” in Proc. the Springer, 2011, pp. 645–675.
2007 ACM Conference on Recommender
Systems, ACM, 2007, pp. 17–24.
[20] F. Peleja, P. Dias, and J. Magalha˜es,
“A regularized recommendation algorithm
with probabilistic sentiment-ratings,” in
Proc. 12th International Conference on
Data Mining Workshops (ICDMW),
IEEE, 2012, pp. 701–708.
[21] S. G. Esparza, M. P. O’Mahony, and
B. Smyth, “Mining the real time web: A
novel approach to product
recommendation,” Knowledge-Based
Systems, vol. 29, pp. 3–11, 2012.
[22] S. M. Kywe, E.-P. Lim, and F. Zhu,
“A survey of recommender systems
in twitter,” Social Informatics, Springer,
2012, pp. 420–433. [23] B. Hajian and T.
White, “Modelling influence in a social
network: Metrics and evaluation,” in Proc.
Third International Conference on
Privacy, Security, Risk and Trust (Passat),
IEEE, 2011, pp. 497–500.
[24] I. H. Witten and E. Frank, Data
Mining: Practical Machine Learning
Tools and Techniques, Morgan Kaufmann,
2005.
[25] X. Amatriain, A. Jaimes, N. Oliver,
and J. M. Pujol, “Data mining

21 https://sites.google.com/site/ijcsis/
ISSN 1947-5500