Advanced Computational Intelligence: An International Journal (ACII), Vol.2, No.

3, July 2015

TRUST METRICS IN RECOMMENDER SYSTEMS: A

SURVEY
MortezaGhorbani Moghaddam1,Norwati Mustapha1, Aida
Mustapha1NurfadhlinaMohd Sharef1, Anousheh Elahian2
1

Department of Computer Science, University Putra Malaysia, Selangor, Malaysia

Department of Information Technology,Virtual University of Shiraz, Shiraz, Iran

ABSTRACT
Information overload is a new challenge in e-commerce sites. The problem refers to the fast growing of
information that lead following the information flow in real world be impossible. Recommender systems, as
the most successful application of information filtering, help users to find items of their interest from huge
datasets. Collaborative filtering, as the most successful technique for recommendation, utilises social
behaviours of users to detect their interests. Traditional challenges of Collaborative filtering, such as cold
start, sparcity problem, accuracy and malicious attacks, derived researchers to use new metadata to
improve accuracy of recommenders and solve the traditional problems. Trust based recommender systems
focus on trustworthy value on relation among users to make more reliable and accurate recommends. In
this paper our focus is on trust based approach and discuss about the process of making recommendation
in these method. Furthermore, we review different proposed trust metrics, as the most important step in this
process.

KEYWORDS
Recommender Systems, Trust metric, Collaborative Filtering, Information overload, E-Commerce.

1. INTRODUCTION
Growth of the Internet has made possible for faster access to information as compared to the past.
But the growing rate of information is very high. The digital data created reached 4 zettabytes in
2013 [1] and the future grow is staggering. Intel predicted that by 2015, it is the numbers of
networked devices across the world will be two times higher than the global population. At that
time, it would take 5 years to view all crossing IP networks at each second [2]. By 2017 it is
predicted that mobile traffic will have grown 13 times in compare with 2012. At that time, there
will be 3 times more connected devices than people on earth [1]. All of these data present that in
the future it is impossible to physically follow the information flow in real time. This is
essentially the phenomena of information overload, whereby there exist a huge amount of
information over the Internet but finding the exact information of interest would be difficult.
While the access to the information may be faster, it is not necessarily easier because finding the
information ofinterest in the huge datasets is time consume and difficult.
To solve the problem, information should be filtered. One of the most important techniques of
filtering of information is recommender systems (RS). RS have proposed to help and guide users
to find items (e.g., sites, books, movies, news, music, etc.) of their interest from a plethora of
available choices in huge pool of data such as the Internet [3]. RS are being used widely in
modern e-commerce applications to cope with information overload problem, however they
suffer from several inherent issues such as data sparsity, cold start users, low accuracy and
malicious attacks.
DOI:10.5121/acii.2015.2301

Advanced Computational Intelligence: An International Journal (ACII), Vol.2, No.3, July 2015

As RS utilize behavior of users to make prediction about their interests, when there is not enough
information about previous behavior of users the problem has been called cold start and causes
the system not be able to make prediction properly. Sparsity problem refers to huge number of
users and items in an e-commerce site. For example in Amazon.com has thousands users who are
able to buy millions items. At this case, for each user, the ratio of the system about previous items
of interest to whole number of items is very small. Evan for active users, with hundreds previous
transactions, this ratio is not comparable. Sparcity problem causes that finding users with
common tastes in a big dataset be difficult.
By introducing Web2.0 websites, which allow their user to interact to each other, Social-based
recommender systems (SRS) as next generation of recommender systems have been proposed.
SRS use relationship between users and social activities metadata to solve challenges of
traditional RS and improve accuracy of recommendations[4]. Trust-based RS use trustworthy
value in the process of making recommendations to improve accuracy of recommendation and
solve problem of malicious attacks. However several proposed trust-based approach have targeted
cold start and sparcity problems and achievements have been followed[3], [5][7].
Trust in RS is defined as ones belief toward others in providing accurate ratings relative to the
preferences of the active user. To measure trust intensity between users, variety trust metric have
been proposed.Some approaches have used explicit trust [8], [9], however others inferred implicit
trust [10], [11].Furthermore, the calculated trust intensity in trust metrics may beglobal or local.In
this paper we describe in detail, the process of making recommendation in Trust-based RS and
focus on trust metrics as the most important step in the process. Furthermore, we have reviewed
the most important trust-based approaches and have summarized their trust metrics.
The rest of this paper is organized as follows: properties of trust and process of making
recommendation in trust-based RS are discussed in section 2. Section 3 describes the most
important trust-based approach and review in detail the trust metrics in these approaches. Finally,
the last section summarizes the paper.

2. THE PROCESS OF TRUST-BASED RECOMMENDATION

Recommender systems can generally be categorized into four types, which are Collaborative
Filtering, content-based, demographic-based, and hybrid approaches. CF techniques are the most
successful techniques for recommendation [3], [12]. This technique capitalizes on history of
users activities while disregarding the actual content of items. This technique requires users to
indicate preferences in the form of ratings. Based on previous ratings, the system recommends
items to user based on its understanding on users interests and finding users with similar tastes
[13].
In content-based methods, the system utilises the previous items of interest to understand the
contents that are favourites of the user. These methods extract certain contents from items of
interest and build relations between users and the extracted contents. Unlike collaborative
filtering, which is not related to items, in content-based methods access to content of items is
essential [14].
Demographic methods use similarity between users to make recommendation. In this respect,
they are similar to collaborative filtering. The key difference between these methods is similarity
metric. Collaborative filtering methods measure similarity between users based on their previous
actions, however demographic-based approaches set demographic information of users as metrics
to measure similarity.
2

Advanced Computational Intelligence: An International Journal (ACII), Vol.2, No.3, July 2015

Usually trust-based recommender systems place in the first category, where the trust intensity is
used as an importance factor to measure the most similar users and aggregating their tastes for
making prediction or recommendation. Later in this section, we discuss about the process of
making recommendation in trust-based RS.
There are four steps in the process of trust-based recommendation. These steps are shown in
Figure 1. The first step, that is the most important step and our focus in this paper, is measuring
the trust. The output of this step is a || || matrix, where || is number of users and content
of cell u, v presents the trust intensity between users u and v. Based on characteristics of trust this
number may be binary or floating point, positive or negative. In some approaches such as [8], [9],
[15][17] the trust value between some users are defined explicitly. At this case the system fills
corresponding cells with explicit values and calculates trust intensity for other cells. On the other
hand, there are implicit trust approaches, which infer value of all cells of the matrix. In these
approaches, the trust value between users is not defined explicitly by users and the system infers
them based on social behavior of users. Although accuracy of explicit trust is more than implicit
one, when explicit trust is not exists, using of implicit trust-based approaches in comparison with
traditional RS can improve accuracy of recommendation results[10], [11].
The trust matrix is sparce. It means that many cells in the matrix are empty. To solve this problem
usually trust propagation methods are used. By trust propagation methods, the system calculate
the trust intensity between users based on non-empty cells of the trust matrix. Be note that there is
different between implicit inferring trust and trust propagation. For inferring implicit trust, the
system focuses on social behavior of users, however for propagating of trust, it focuses on
previous calculated trust values. For this reason usually trust relationship between users present as
a graph, called trust graph. Trust graph is a two dimensional graph, that is constructed based on
the trust matrix. In this graph users are presented as nodes and trust value between users are
presented as edges. Furthermore, the trust intensity between users is used as weight of the edge. If
all cells of the trust matrix are non-empty therefore the graph is complete. It means that trust
intensity between all users are calculated. But most of time the graph is not complete, even it may
not be connected. Trust propagationis one of the most important sub processesthat calculates trust
intensity on missed edges.
Neighbors are the most related users to the active users, that the system utilizes their information
to make prediction about interests of the active users. Similar to traditional collaborative filtering
approach, the second step in trust-based approaches is selecting the neighbors. But there is an
important difference here. In traditional approaches neighbors are chosen based on similarity,
however in trust-based approaches the trust value is key parameter.
One intrinsic problem of collaborative filtering approaches is malicious attacks. This problem
refers to this step because neighbors can affect recommendation results. In collaborative filtering
approaches, the system measures similarity between users based on previous common tastes. If an
attacker copy activities of the active users, therefore the system choose him/her as the most
similar users. It means that he/she is able to drive recommendations by using his next rates. To
solve this problem trust-based approaches uses calculated trust value between users to choose
more reliable users as neighbors. Previous researches presented that user with high trust intensity

Trust
calculation

Trust
Propagation

Neighborhood
selection

Aggregation

Presentation

Figure 1. Steps of recommendation in trust-based approaches

3

Advanced Computational Intelligence: An International Journal (ACII), Vol.2, No.3, July 2015

may have low similar tastes[3], [18]. Therefore usually combination of similarity and trust is used
in selecting neighbors to improve accuracy of recommendation and solve malicious attacks [3],
[5], [6].
The third step is aggregating tastes of neighbors. At this time the system predicts interest of the
active user u about non-rates item i. for this reason it chooses the neighbors who have rated the
items i previously and aggregates their rates:
 ,
=    ,

(1)

Where  and are predicted and actual rates respectively. One of the most common aggregation
formula, is Resnick[19] that is formulate as equation 2.
 ,
=  +

,

$, 

|$, 
|

(2)

Where  is average previous rates of user v and &, 

is weight that presents importance of v on
us point of view. As discussed before, in trust-based approach & is combination of similarity and
trust intensity.
The final step in making recommendation is showing the results. This step create output results
that may be: prediction, recommendation or ranked list. In prediction, the input is the active user
and the active item. At this case the system should predicts interest of the user about the item. But
in recommendation, the input is only the active user. At this case the output is a list of items that
are more related to the user. The number of items in the list may be define as an input threshold.
In other hand, the input threshold is minimum interest value. At this case the system predicts
interest of the user to all items and shows only the items that there  is greater than the input
threshold. If output result of recommendation is a sorted list of items based on predicted rates, the
output is a ranked list of items.
3. TRUST METRICS
As discussed in previous section, the method of calculating and propagating trust is very
important and affects result of recommendation. In this section we review in detail the different
trust metrics in trust-based approaches. Also main contribution of discussed approaches are
summarize in Table 1 at the end of this section.
Advogato[20]has been proposed by Levin in order to discover the most trustworthy users. The
main goal of Advogatos has been attack resistance. It does not make prediction or
recommendation and just find the most trustworthy users however there is no distinction between
the trusted users in this approach. Therefore it is not appropriate for recommendation. Advogato
transforms the trust network into a flow network. This transformation is the essence of the
algorithm. The computation of the trust metric is performed relative to a seeds of trustworthy
users, a limited count of users who explicitly have defined as base users. Advogato uses a breathfirst search algorithm from the seed to assign capacities to nodes. For this reason it uses the
distance from the seeds (' ) and the out-degree of nodes with a smaller distance (if any).
()* +,-. 
= '

(3)

Ziegler proposed Appleseed [21] in his Ph.D. thesis. His approach is similar to Advogato. But
here, edges in trust graph are assigned continuous weights. In contrast to Advogato, Appleseed
uses is suitable for distributed implementation [22]. It models trust as energy and propagates it
4

Advanced Computational Intelligence: An International Journal (ACII), Vol.2, No.3, July 2015

using spreading activation mechanism. The output of the algorithm is the top-N nearest trust
neighbors for every user. The novelty of this approach lies in the propagation step, whereby a
source node u is activated through an injection of energy. The energy then propagates to other
nodes through edges of the graph. The ratio of energy propagates on each edge is related to the
weight of the edge. Therefore, the more direct trusted users in the network, receive higher level of
energy. AppleSeed supposed that trust is additive. Therefore, the amount of energy received by a
user is the sum of total energy that travels through variety of paths. If there are many weakly
trusted paths between two users, this pair of users will obtain a high trust value. Equation 4
presents trust metric in Appleseed. Where /
is set of nodes who there is an input link from
them to u. 0/  , 
presents the propagated energy from v to u when the source energy injected
in node s.
()* +1123, ), 
=

 

0/  , 

(4)

Golbeck proposed TidalTrust[8] as a recommender system that used explicit trust to make move
prediction. Like Appleseed, main contribution of TidalTrust is related to trust propagation, where
it only considers the paths of the shortest length, disregarding others. Among the shortest paths,
TidalTrust calculates the maximum of the minimum edge weights on a path, Smaxmin. The
minimum of the edge weights on a path is therefore considered as a measure of weight of this
path, as in Advogato. After the path selection process, TidalTrust again uses multiplication to
calculate the weight of a path from its edge weights, as in Appleseed. Smaxminis used as a
threshold value to select the best paths. The paths that will be considered in the final trust
computation are only the shortest paths from with a minimum of the edge weights that is not less
than Smaxmin. Let ( 3 be set of trustworthy users of s whom the trust value is greater than or
equal to the Smaxmin, and ()*5612. ), 
be explicit trust between s and v. Therefore
Golbeck used below recursive formula for trust propagation:
()* 7,-27. ), 
=

879()*5612. ), 
()* 7,-27. , 

87 9()*5612. ), 

(5)

Kuter and Golbeck proposed SUNNY [16] as a trust-based recommender system. The process of
propagation trust in SUNNY is similar to TidalTrust. The key difference lies in the selection of
paths that contain specially selected neighbours of destination users. In SUNNY each neighbour
of destination node B has one of three states: unknown, include or exclude from the set of
paths used for the final trust calculation. In the beginning, the Bayesian network is used to
calculate bounds on the confidence of S in Bs recommendation when all neighbours of B have an
unknown state. After this initial step, Sunny iterates through all neighbours of B and tries to
change their state to include and again calculates confidence bounds for the belief of S in a
recommendation from B. If the resulting confidence bounds are too different from the original
confidence bounds, then the state of this neighbour of B is changed to exclude. Finally, only the
neighbors of B that do not change the original confidence bounds too much are selected.
For this reason, SUNNY calculates differences between users at three levels: Overall difference,
Difference on extremes and Maximum difference. It also calculates confidence between users
based on Pearson Correlation Coefficient (PCC) similarity metric. To estimate trustworthy from
source node S to destination node B, it performs a variant of backward breadth-first search over
the nodes of the trust network towards the source node S and constructed a Bayesian Network by
using the trust network, S and B. SUNNY computes conditional probabilities for every edge in
the Bayesian Network based on computed confidences and levels of differences between users.

Advanced Computational Intelligence: An International Journal (ACII), Vol.2, No.3, July 2015

Massa and Avesani proposed MoleTrust[23] as a trust metric that is similar to TidalTrust in the
sense that it also uses explicit trust. The major difference lies in the propagation method. While
TidalTrust used depth-first search for finding the highest trust users, MoleTrust used breadth-first
search. In this approach, a maximum-depth is specified as an input and all raters will be
considered up to the maximum-depth.
MoleTrust includes two steps: Purpose of the first step is to destroy cycles in the graph. Second
step consists of a graph walk starting from source node with the goal of computing the trust value
of visited nodes. For this reason the initial trust value of the source user is set to 1 and to compute
the trust value between u and v, MoleTrust performs a backward exploration, whereby the trust
value from u to v is the aggregation of trust values between u and users directly trusting v is
weighted by the direct trust values.
()*:27. ), 
=

81,()*5612. , 
()*:27. ), 

81,()*:27. ), 

(6)

Here, intensity of trust is depend on predicted trust for users in previous level (;0<0=0))>)

and trust statement (()*5612. ) between them to the node u. For predicting the trust value of a
user, MoleTrust analyses all the incoming trust edges and accepts only the ones coming from
users with a predicted trust score greater or equal than a certain threshold.
A Trust-aware Recommender System (TaRS) is also proposed by authors of ModelTrust in [17].
The trust metric in TaRS is same as MoleTrust and difference lie in aggregation step, where TaRS
uses combination of trust and similarity as weight.
In [11], trust is defined and used differently whereby the trust value is represented in terms of
percentage of correct predictions. This research proposed a new metric for inferring global
implicit trust by distinguishing two types of trust for each user. The first type of trust is item-level
trust, which is measured by perceived trustworthiness of the user to each item. Item-level trust is
calculated based on ratio of correct predictions of the user to all of his/her predictions about the
item. Suppose that ?0< =*),
is set of all prediction of user u about i. some of these
prediction have been correct and other have been incorrect. If ?0< =*) @. ,
be set of
correct predictions, at this case item-level trust is formulated as below equation:
()*A.B ,
=

|?0< =*) @. ,

|
|?0< =*),
|

(7)

ODonovan and Smyth proposed also profile-level trust similar to item-level trust: For each user,
ratio of correct predictions to all predictions, independent of items, introduced as profile-level
trust.
|?0< =*) @. 
|
(8)
()* 1C2 ,
=
|?0< =*)
|

Experiments in [11] presented that accuracy of item-level is 22% more than collaborative
filtering. However this value for profile-level trust is around 15%. Furthermore experiments in
[11] showed that using of combination of trust and similarity in aggregation step, improves
accuracy of results.
Zhang and Xu proposed Topic-level trust in [10]. This method is similar to item-level trust [11] in
the sense that it also uses correct prediction to measure implicit and global trust. But it uses topics
in taxonomy to calculate trustworthiness.
6

Advanced Computational Intelligence: An International Journal (ACII), Vol.2, No.3, July 2015

The trust metric in [10] is formulated as below equation, where ()* 71 , *
denotes trust
intensity of user u about topic t, ()* .B ,
denotes items-level trust and D*0E)*
is set of
items that are related to topic t.
(9)
A.B.
()*A.B ,

()* 71 , *

=
|D*0E)*
|
Zhang and Xu also proposed new method to calculate item-level trust. by. They used ratio of sum
of trust value to number of raters (Equation 10):
| ,,
G,
|
(10)
-.
1

3
()*A.B ,
=
|HI*0)
|
Where  , , 
is the predicted rate by v about interest of u to i, S represents the rating scale, and
HI*0)
denotes set of users, except u, who have rated item i.Experiments in [10] showed that
topic-level trust produces lower error (around 13%) than traditional RS.
Jamali and Ester proposed TrustWalker[6] as a random walk model that combines trust-based and
item-based CF approaches to further improve prediction accuracy and solve cold-start as well as
sparsity problems. This method uses explicit friendship to calculate local trust.Like Appleseed,
the trust metric in Trustwalker is related to out-degree of nodes, but here all edges have equal
weight. Suppose that there is an edge between u and v, thus TrustWalker calculates trust value
between the nodes by using below equation:
()* 7.J-2 , 
= 1|>*
|

(11)

Where >*
is set direct friends of u. The novelty of TidalTrustis on the proposed random walk
model.In this approach, to predict the rate of user u about item i, the system uses direct friends. If
they have rated item i before, the system returns the rate. Otherwise, the system returns one of
previous rated items based on probabilistic calculation or asks their direct friends. The proposed
probabilistic model in [6] is related to maximum similarity of items that are being rated by current
user.
Given that size of the set of common users is also important Jamali and Ester proposed a new
similarity metric to affectLM,N L, as number of common users who have rated to both items, in
computing similarity between users. They used a sigmoid function based on Pearson Correlation
Coefficient (SPCC) to avoid favoring the size of common users too much:
) E , O
3P@@ = ) E , O
P@@ .

1 + exp 

L@UV,W L
X

(12)

If the size of the set of common users is big enough, then the second part of the equation would
converge to 1, but for small sets of common users, the factor would be 0.6. The number 2 in the
denominator of the exponent is because they wanted to have a factor of greater than .9 if the size
is greater than 5. Based on experiments in [6], Jamali and Ester claimed that considering the size
of the set of common users in the item similarity metric reduces the error.
The experiments on Epinions data set showed especially for cold-start users, who have only a few
ratings, TrustWalker achieved better results in comparison with other methods. For example in
comparison with traditional collaborative filtering, it improved accuracy of recommendation
around 20%, while it increased the coverage more than three times. Jamali and Ester compared
their proposed approach with other trust-based approaches, too. The result presented that its
improvement on accuracy was around 2% and 17% for TidalTrust and MoleTrust respectively,
while the improvement on coverage was more than 20% for both compared approaches.
7

Advanced Computational Intelligence: An International Journal (ACII), Vol.2, No.3, July 2015

A modified version of TrustWalker, called MWalker[5], was proposed by Jin and Chen.
MWalker used multiplication of binary friendship, as explicit trust, and similarity between users
to calculate trust value.
()*:J-2 , 
= ()*5612. , 
Y E7- , 

(13)

MWalker used a breadth-first search method to make recommendation. For this reason it starts
from source node u and sorts his/her friends according to the trust value between them. Top-K
most trustworthy users compose the candidate set. The algorithm visits each user in the candidate
set sequentially. If the user rated to the target item i, then returns it directly. Otherwise the user
returns his/her rate about item j that is the most similar item to i. The described algorithm
continues for all candidate users till maximum defined depth.
SimTrust[24] has been proposed by Bhuiyan et al. as a recommender system. In SimTrust, trust
has been defined and used in a different way. In this method, the implicit local trust is inferred
using tags based on the assumption that a trustworthy text description about each item exists.
Therefore, to solve ambiguity problem of tagging, the system is able to extract the semantic
meaning of a tag based on the description of the items. Bhuiyan et al. used tf-idf as a text mining
approach to extract frequent keywords from description of items. Suppose that Wut is set of
frequent keywords about tag t that is used by user u. Actually Wut represents the semantic
meaning of the tag t as point of view of user u. SimTrust measured trust value between users by
using following equations:

/
()* 3B7. , 
= Z 4  [\|$|
(14)
/
J


Where /
denotes number of tags in intersection of Wu and Wv that contain keyword k. W is set
of all keywords in Wu or Wv.

Bhuiyan et al. evaluated their approach by using Amazon book data set for different
neighborhood size. Result showed increasing the neighborhood size, improve the
recommendation results. They also compared SimTrust with TidalTrust and traditional
collaborative filtering. The result presented that SimTrust performed better that compared
approaches. It is concluded that the method is helpful in dealing with data sparsity problem.
Guo et al. proposed Merge [3] as a trust-based method to solve cold-start and sparsity problems.
Instead of using direct rated items for prediction, Merge uses extended rated items. Extended
items are the items being rated by the user explicitly or by at least one of his/her direct friends.
Suppose that item i is a member of the extended set for active user u. If the user u rated directly
the item i, therefore his/her previous rate will be saved. Otherwise, aggregation of rates of direct
friends will be used as the predicted rate for the item in extended list. We call these type of rates,
which are aggregated based on rate of direct friends, merged-rates.
The trust metric in Merge is same as MoleTrust. But the contribution of Merge is in choosing
aggregation section, where it uses a linear combination of trust, rating similarity and social
similarity as importance weight:
$, 
= ]^ Y E-. , 
_ + ]` ()*:27. , 
_ + 1 ^ `
Y E 3-2 , 

(15)

The experimental results on three real-world datasets, FilmTrust, Flixer and Epinion, showed a
significant improvement against other methods in terms of both accuracy and coverage, as well as
the overall performance. In addition, this approach has performed well in cold-start problem.
8

Advanced Computational Intelligence: An International Journal (ACII), Vol.2, No.3, July 2015

Ray and Mahanti focused on trust propagation. They analyzed Epinion dataset and observed that
majority of the trust statements passed between users cannot be seen as a reflection of similarity
between the two users. They proposed RN [18] as a trust-based approach, to reconstruct the trust
networks in order to improve the prediction accuracy. This is done by using combination of trust
and similarity between users as the weight of trust in trustworthy graph. The approach then
pruned the graph based on a threshold correlation value. They removed all edges in the original
trustworthy graph is the correlation value between two users is less than the defined threshold.
This approach constructs the network at different propagation levels. For this reason it uses a
linear decay approach for propagating trust. The formula used for the trust metric is < / +
1
/< where d is the maximum propagation distance and n the distance of the target user from the
source user.
Experiment presented that reconstructing the trust network improve accuracy of recommendation
in comparison with complete trust network. However RN achieved poor coverage and it failed to
function in cold conditions where user similarity may not be computable.
A Trust-based Ant Recommender System (TARS) that uses dynamic trust for prediction has been
proposed in [25] by Bedi and Sharma. TARS focuses on sparsity and cold-start problems and uses
trust graph as well as biological metaphor of ant colonies to choose the best neighbors. In order to
measure initial trust value (()*b
between users, TARS uses a combination of similarity and
confident as below equation:
2 ) E, 
=>/g|

g ) E, 
0I/< =>/g|
0
) E, 
+ =>/g|

k
()*b , 
=
j =>/g|
g ) E, 
= 0I/< =>/g|
0
d
c
0 g ) E, 
= 0I/< =>/g|
= 0
e

(16)

Where =>/g|
is amount of confidence a user u should have on user v, which is calculated as
ratio of number of co-rated items.
In the second phase, TARS uses the users constructed trust network with several ants that are
placed in root of the network at initialization time. The ants move around the network and find
the most trustworthy users. To selecting neighbors, they use trust valueson edges, as pheromone,
and level of connecters from source node to transfer.
After aggregating, which is calculated by Resnicks formula, updating the pheromone values is
final step. In TARS the trust value at time t, is calculated as below equation:
()*. , 
= 1 l
()*.Gm , 
+

pq
(_*I=0< rm
()*.Gm  , 

(_/I*0< <

(17)

Where l is evaporation rate, < is the level of connectedness from the source node S to
v,T_tracedv denotes the number of items traced by user v which are unrated by S, andT_unrateds
denotes the total number of items unrated by S.
Authors of TARS, experimented their proposed approach by using two real datasets, Jester and
MoveLenz. As it was expected, the result presented that increasing the number of neighbors,
decreases the average quality of the recommendations. Also comparison between TARS and
traditional collaborative filtering showed, TARS overcomes traditional CF especially over time,
when it updates trust value. Furthermore, TARS identified popular users with high DTP as default
trusted friends. It claimed thatto solve cold-start problem, the list of default trusted friends can be
captured and automatically added to the new users list at initial stage.
9

Advanced Computational Intelligence: An International Journal (ACII), Vol.2, No.3, July 2015

AgeTrust[26]focused on aggregation method and proposed a temporal weighting method. This

research claimed that older friends are more reliable than new one. Therefore they used age of
friendship to rank trusted users. It means that in AgeTrust, the importance weight of older trusted
usersaremore than newer ones. Suppose that user u has n friends and position of user v in the rank
list, based on time of friendship, is p. At this case the importance weigh is calculated as bellow
equation:
$, 
= Y E, 
1 s; u

1
vw ()*, 

/1

(18)

Where ^x[0,1]is a minimum trust value which is assigned to newest friend, n denotes number of
direct friends of user u. To validate the proposed approach, [26] used Delicious dataset and
calculated preference of user to items based on number of assigned tags. Comparison against the
traditional collaborative and pure friendship-based approaches showed that the prediction
accuracy of AgeTrust, is more than other approaches. It means there is a positive relation between
the age of trust and the trust value.
Main contributions of discussed approaches in this section are summarize in Table 1.
Table 1. Main contribution of trust-based approaches
Trust measurement
Trust-based
approaches

Trust
calculation

Trust propagation

Advogato

AppleSeed

Item-level trust

Profile-level trust

TidalTrust

SUNNY

MoleTrust

Neighborhood
selection

TaRS
Topic-level trust

TrustWalker
SimTrust
RN
TARS

AgeTrust

MWalker
Merge

Aggregation

10

Advanced Computational Intelligence: An International Journal (ACII), Vol.2, No.3, July 2015

4. SUMMARY
Traditional recommender systems suffer from several inherent issues such as data sparsity, cold
start users, low accuracy and malicious attacks. Social recommender system use metadata such as
relation among users and social behaviour of users to solve these problem. Trust-based
recommender systems are the most successful approaches in this generation. They utilizes trust to
create more accruable and reliable predictions. Trust in recommender systems is defined as ones
belief toward others in providing accurate ratings relative to the preferences of the active user.
In this paper we discussed on process of making recommendation in trust-based approaches. We
categorized this process to four steps: Trust measurement, selecting the neighbours, aggregation
and presentation. Furthermore, we reviewed in detail variety trust-based approaches and argued
about different steps in these methods. Also, the main characteristics of trust, such as globality
and visibility have been discussed in different approaches. Aware of this review, the most
approaches have used explicit trust rather than implicit one. It may be due to accuracy of explicit
trust that is more than implicit one. However to solve sparcity problem, propagating of trust is
effective and imperative. Therefore different approaches used various propagation methods.
Based on reviewed methods, accuracy and coverage of trust-based approaches is better than
traditional recommender system. Furthermore, these methods achieved significant improvements
about cold start data, when there is not enough information about behaviour of users. It seems that
focus of first trust-based approaches were more on trust metrics, however recent approaches are
more related to selection of neighbours and aggregation.
REFERENCES
[1]

Internet-minute-infographic. (2014) What Happens in an Internet Minute?. Retrievef from August 3,

2014, http://www.intel.com/content/www/us/en/communications/internet-minute-infographic.html. .
[2] Internet-minute-infographic. (2011) What Happens in an Internet Minute?. Retrievef from January
23,
2012,
http://www.intel.com/content/www/us/en/communications/internet-minuteinfographic.html. .
[3] G. Guo, J. Zhang, and D. Thalmann, Merging trust in collaborative filtering to alleviate data sparsity
and cold start, Knowledge-Based Syst, Dec. 2013.
[4] N. Ghorbani Moghaddam, Morteza Mustapha, A. Mustapha, N. Mohd Sharef, and A. Elahian, A
Review on Trust-based Recommender Systems.
[5] J. Jin and Q. Chen, A trust-based Top-K recommender system using social tagging network, in
Proceedings of the 9th International Conference on Fuzzy Systems and Knowledge Discovery (FSKD
2012), 2012, no. Fskd, pp. 12701274.
[6] M. Jamali and M. Ester, TrustWalker: a random walk model for combining trust-based and itembased recommendation, in Proceedings of the 15th ACM SIGKDD international conference on
Knowledge discovery and data mining, 2009, pp. 397406.
[7] G. Guo, J. Zhang, and D. Thalmann, A Simple but Effective Method to Incorporate Trusted
Neighbors in Recommender Systems, in 20th International Conference on User Modeling,
Adaptation and Personalization (UMAP 2012), 2012, no. Imi, pp. 114125.
[8] J. Golbeck, Generating predictive movie recommendations from trust in social networks, in
proceedings of the 4th international conference on Trust Managementi (Trust06 ), 2006, pp. 93104.
[9] P. Massa and P. Avesani, Trust metrics on controversial users: Balancing between tyranny of the
majority, Int J Semant Web Inf Syst, vol. 3, no. 1, pp. 121, 2007.
[10] Z. Fu-guo and X. Sheng-hua, Topic-level Trust in Recommender Systems, in 2007 International
Conference on Management Science and Engineering, 2007, pp. 156161.
[11] J. ODonovan and B. Smyth, Trust in recommender systems, in Proceedings of the 10th
international conference on Intelligent user interfaces - IUI 05, 2005, p. 167.
[12] C. Li, L. Ma, and K. Dong, Collaborative Filtering Cold-Start Recommendation Based on Dynamic
Browsing Tree Model in E-commerce, in 2009 International Conference on Web Information
Systems and Mining Collaborative, 2009, no. 2, pp. 620624.
11

Advanced Computational Intelligence: An International Journal (ACII), Vol.2, No.3, July 2015
[13] M. Ghorbani Moghaddam, N. Mustapha, A. Mustapha, N. Mohd Sharef, and A. Elahian, A Review
on Similarity Measurement Methods in Trust-based Recommender Systems, in 9th International
Conference on e-Commerce with focus on E-Trust (ECDC 2014), 2014.
[14] M. Qin, Q. Yang, and F. Fu, A context-aware media content personalized recommendation for
community networks, 2013 Int Conf Wirel Commun Signal Process, no. 2, pp. 16, Oct. 2013.
[15] J. Golbeck, Personalizing applications through integration of inferred trust values in semantic webbased social networks, in Proceedings of Semantic Network Analysis Workshop, 2005, pp. 1528.
[16] U. Kuter and J. Golbeck, Sunny: A new algorithm for trust inference in social networks using
probabilistic confidence models, in AAAI07 Proceedings of the 22nd national conference on
Artificial intelligence, 2007, vol. 2, pp. 13771382.
[17] P. Massa and P. Avesani, Trust-aware recommender systems, in Proceedings of the 2007 ACM
conference on Recommender systems, 2007, pp. 1724.
[18] S. Ray and A. Mahanti, Improving Prediction Accuracy in Trust-Aware Recommender Systems, in
the 43rd Hawaii International Conference on System Sciences, 2010, pp. 19.
[19] P. Resnick, N. Iacovou, M. Suchak, P. Bergstrom, and J. Riedl, GroupLens: an open architecture for
collaborative filtering of netnews, in Proceedings of the 1994 ACM conference on Computer
supported cooperative work - CSCW 94, 1994, pp. 175186.
[20] R. Levien, Attack resistant trust metrics, UC Berkeley, 2004.
[21] C. Ziegler, Towards decentralized recommender systems., 2005.
[22] A. Wierzbicki, Trust and Fairness in Open, Distributed Systems. 2010, p. 244.
[23] P. Massa and P. Avesani, Trust metrics on controversial users: Balancing between tyranny of the
majority, Int J Semant Web Inf Syst, vol. 3, no. 1, pp. 121, 2007.
[24] T. Bhuiyan, Y. Xu, A. Jsang, H. Liang, and C. Cox, Developing trust networks based on user
tagging information for recommendation making, in Web Information Systems Engineering WISE
2010 LNCS, 2010, pp. 357364.
[25] P. Bedi and R. Sharma, Trust based recommender system using ant colony for trust computation,
Expert Syst Appl, vol. 39, no. 1, pp. 11831190, Jan. 2012.
[26] M. G. Moghaddam, N. Mustapha, A. Mustapha, N. Mohd Sharef, and A. Elahian, AgeTrust: A new
temporal trust-based collaborative filtering approach- In press, in the proceeding of 5th International
Conference on Information Science & Applications (ICISA 14 ), 2014, pp. 14.

12