There it can improve the accuracy of recommendations. Due

There has been an increasing interest of
researchers in the area of trust based recommendations. Using trust as
additional social information source provides another view of user preferences
other than item ratings and it can improve the accuracy of recommendations. Due
to the inability of standard CF approaches to find sufficient similar neighbors
in sparse data sets, many researchers have been start studying using the trust
in order to generate more personalized recommendations, in which the users
receive recommendations for items rated highly by people in their web of trust
(WOT), or even by people who are trusted by these WOT members. X. Chen et.al 22 propose a new trust based on collaborative algorithm
in social networks using the Dijkstra algorithm in measuring the trust between
users as additional information source to improve the recommendation systems.
The proposed approach starts with finding neighbors by measuring similarity
between users using Pearson Correlation metric. Then, the user’s relationships
are represented as undirected weighted graph, where the weight of each edge
represents the direct trust between two users. After the relationships graph
was constructed, the Dijkstra algorithm will be used to find the shortest path
as the ultimate trust between target user and other users in the network. Using
Flixster dataset, the item is predict based on a weighted sum of ratings given
by trusted neighbors to item. However, this approach may not yield a realistic
result since the trust relation represent in undirected graph and doesn’t take
into consideration the asymmetry property of the trust, where the trust depends
from user to another, it is personal, subjective and may hold various opinions.  For example, considering two users u and v,
user u trusts user v does not mean that user v trusts user
u. The
authors of 23 modeled user relationships based on the “Six Degrees
of Separation” social theory, in order to evaluate the trust between users. In
this paper, the trust is evaluated based on the minimum number of layers linked
between target user and other users in the network. The top-K trusted users for
the target user are identified and the prediction is generated by averaging
their ratings weighted by their trust value. This technique suffers from a
scalability problem and cannot deal with new coming user who has not have any
relationship yet. In paper 14, Y. Guo et.al,
propose computational model of trust and a predictive algorithm based on the
trust and similarity measure. The main hypothesis of the authors was: If a user
makes more recommendations for others, other users will have greater trust in
him. Based on this assumption, the trust parameter defined as trusted degree by
others in the CF system. The trust degree computed based on two factors. First,
is the trust level of target user being trusted by other users in the
recommendation system. The second factor is the times of user making
recommendation for others. Based on these factors, the trust value between
users is calculated and used in predicating item rating in addition with
similarity. In this paper, instead of measuring the trust value between
particular user and other users, the trust is measured for each user
individually based on his own behavior without considering relationships
between other users and the target user. In this situation, the recommendation
may not reflect the real user trustworthy group. In addition, cannot provide a
recommendation for new user, who have not performed any recommendation before. A
trust model based on Beta model is adopted in 11 to calculate the trust values from the target user’s
point of view. The aim is to determine user’s trustworthiness on recommending
items to the target user, the trustworthy users are distinct from untrustworthy
ones using the trust model. In the classic Beta trust model, the user’s
behavior is a binary value as either “good” or” bad”. Following the Beta
distribution in the proposed trust module, they we do not directly adopt binary
values for a user’s behavior due to the multivariate rating values available in
recommender systems. Instead, they quantify a user behavior as a continuous
value in the range of 0, 1 based on the rating offset between the target user
and the other users. The trust value is computed based on the user’s behavior,
then neighbors with high trust values are selected in order to performs the
prediction. P. Moradi and S. Ahmadian ,17  propose a Reliability-based Trust-aware Collaborative
Filtering (RTCF) method to improve the accuracy of the trust-aware recommender
systems. The trust network for the active users is first constructed based on
the Pearson correlation coefficient measure as final similarity values. This
network is a weighted directed graph where the neighbors of the active user are
nodes of the graph and the adjusted similarity values between two nodes forms
the weights of their associated edge. Using this weighted graph, the trust
between pair of the users is calculated. After measuring the similarity and
trust between users, the trust network can be used to predict the initial rates
of the unseen items for the target user. However, the trust value is based on similarity
as initial step, which makes this method unsuitable when the data are sparsity
and with cold-start users. Parham et.al 24 propose a novel method to determine effectiveness of
the users in trust network of the active user. A game theory called the Pareto
dominance concept is used to identify dominance users of the active user and
the trust statements between users are calculated based on this concept. This
concept is used to identify those trustable users who correctly represent the
interests of the user and who therefore should be considered as candidate
neighbors. The proposed method involves a pre-filtering process that reduces
the effectiveness factor of the least representative users from those of
selected trust users. Moreover, model-based CF can also employ trust metrics in
the development of trust based recommender
systems to improve recommendation performance. TrustSVD25 is one examples of those approaches, it is a trust based
matrix factorization technique. This method incorporating both the explicit and
implicit effectiveness of trusted users on the prediction of items for a target
user.  The explicit influence of trust
values is used to constrain that user specific vectors should conform to their
social trust relationships. Which ensures that user specific vectors can be
learned from their trust information even if a few or no ratings are given. In
this way, the data sparsity can be better alleviated. However, in the trust
rating network, e.g. Epinions, the user can be associated with two different
roles, “trusters” is called for who trust others and “trustees” who are trusted
by others. Based on this observation, Weilong.Y et.al 26propose a model called RoRec to learn dual role
preferences for trust aware recommendation. Different preferences of the two
roles of users are learnt by modeling both explicit and implicit interactions.
Specifically, truster and trustee specific preferences are estimated to fit the
explicit ratings and trust relations using matrix factorization techniques.
Distinct from existing methods which measure two user correlations solely based
on single links between two users, in RoRec, local links structure of trust
network is leveraged to evaluate the correlations between two trusters/trustees
for modeling implicit interactions. Even if those model-based approaches faster
in prediction, avoid over fittings and they scalable to the large and actual
dataset. In addition, it is unable to generate recommendations for those users
who have not provide any ratings before. In addition, the modeling process is
time consuming and may cause information loss, which leads to the drop of
recommendation accuracy