Trust-aware Social Recommender System Design

Peixin Gao

, John S. Baras

and Jennifer Golbeck

Institute for System Research, and Department of Electrical & Computer Engineering, University of Maryland,

College Park, MD, U.S.A.

College of Information Studies, and Department of Computer Science, University of Maryland,

College Park, MD, U.S.A.

1 STAGE OF THE RESEARCH

With the explosive growth of information and the de-

velopment of the Internet, we are overwhelmed with

an extreme load of information. As a result, it is

difﬁcult to make decisions in front of immense va-

riety of options. Recommender systems are thus de-

signed to overcome the problem of information over-

load created by the Internet. However, current ap-

proaches for recommender system still suffer from

the problems such as sparse information, cold start,

and adversary attacks. On the other hand, social net-

work sites (SNS), like Facebook and Epinions, offer a

good source of knowledge for recommendation. The

idea of integrating signals from social network to im-

prove the performance of the recommendation algo-

rithm has been well accepted and has attracted an in-

creasing amount of research in both academia and in-

dustry, and many social network-based recommender

system designs have been proposed and evaluated.

In this work, we develop a trust-aware recom-

mender system. We interpret connections in associ-

ated social graph as trust relationships among users

in the recommender system, and establish a trust net-

work accordingly. Within the trust network, we pro-

pose models for trust propagation and aggregation,

and design trust-aware recommendation algorithms to

predict the preferences of users over items (services).

Specially, we handle indirect trust in our model,

which could enlarge the information source to a large

amount. We also discuss the issue of distrust (i.e.

negative trust, negative opinion) and propose a way

to consider both trust (positive opinion) and dis-

trust in our model. We also consider integrating our

trust-aware recommendation framework with classic

collaborative ﬁltering to take advantage of both ap-

proaches and further improve the performance in rat-

ing estimation and item recommendation. As an ex-

ample of application scenario, such framework of

trust-aware recommender system design can be ap-

plied for directed SNS like Epinions and Delicious.

Currently we are at the stage of evaluating the ac-

curacy and efﬁciency of the system with both syn-

thetic and real data. Meanwhile, we are trying to ex-

tend our model to further improve its performance.

Peixin Gao is the PhD student working on the

topic under the supervision of Dr. John S. Baras and

Dr. Jennifer Golbeck.

2 OUTLINE OF OBJECTIVES

The objectives of our work on trust-aware recom-

mender system design are ﬁve-fold:

1. We discuss the limitations of current recom-

mender systems and the possibility of introducing

information from aligned social networks to im-

prove the performance of recommender systems.

2. We interpret connections (links) in social net-

works as trust relationships among users, and pro-

pose a trust network comprised by users in the so-

cial graph and trust relationships among them, and

develop an appropriate model for trust propaga-

tion and aggregation as well as trust value update

within the trust network.

3. Based on the trust network and trust model, we

design the recommendation algorithms to pre-

dict item ratings of users with information from

trusted neighbors in a collaborative ﬁltering fash-

ion. Specially, we take care of both indirected

trust and distrust (negative trust).

4. we also propose ways to integrate our trust-aware

recommendation algorithms with classic user-

based collaborative ﬁltering method to make the

novel recommendation algorithm capable of ex-

ploiting information from both rating matrix and

trust network to provide better recommendations.

5. We explore directions to further extend our model

for better performance and wider application sce-

narios.

Gao P., S. Baras J. and Golbeck J..

Trust-aware Social Recommender System Design.

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

3 RESEARCH PROBLEM

3.1 Recommender Systems

Recommender system (RS) can be deﬁned as an in-

formation ﬁltering system that is used to predict the

rating or preference that a user would give to an

item in the system (Adomavicius and Tuzhilin, 2005),

and offer different users diverse suggestions. The

goal of recommender systems is to provide person-

alized recommendations of items that suit a user’s

taste, in order to increase the number and variety of

items sold, meanwhile improve user satisfaction and

ﬁdelity. Recommender systems usually have two ap-

plications; one is to predict the extent of interest the

user has in an item which she has not yet rated, the

other is to help the user ﬁnd items that she is interested

in but have not previously encountered. With decades

of development, recommender systems have become

a fundamental and important technology to help peo-

ple make decisions and selections that ﬁt their prefer-

ence from an extreme overload of information, main-

taining the loyalty of the customers and increasing the

sales. Electronic retailers and content providers like

Amazon and Netﬂix are increasingly adopting recom-

mender systems in their platforms to improve user ex-

perience as well as their own proﬁts.

Referring to the taxonomy of (Burke, 2007;

Zanker et al., 2011), the classic approaches in RS

can be classiﬁed into four groups, namely Collabora-

tive Filtering, Content-based, Knowledge-based and

Hybrid Recommendation. While collaborative ﬁlter-

ing exploits item ratings to derive recommendations,

content-based approaches rely on product features

and textual descriptions, and knowledge-based algo-

rithms reason on explicit knowledge models from the

domain. Hybrid ones integrate different approaches

together. Among all these approaches, Collaborative

Filtering (CF) is currently the most successful and

widely implemented. For example, user-based col-

laborative ﬁltering approach (Herlocker et al., 1999;

Koren, 2008) uses Eq. 1 to predict r

, i.e. user u

’s

rating about item o

ˆr

= b

∑

∈S(k;i)

i j

· (r

− b

)

∑

∈S(k;i)

i j

(1)

where S(k;i) is the neighbor set of u

about o

, i.e. the

set of users that are most similar to user u

in terms of

rating history, with s

i j

is the similarity between user

and u

, ∀u

∈ S(k; i). b

and b

are the baseline

estimates for r

and r

respectively, and r

is the

rating of user u

about item o

CF has several advantages in application. First

of all, CF techniques don’t require domain knowl-

edge or extensive data collection (Koren, 2008), and

could scale well to large item bases. In addition,

relying directly on user behavior allows uncovering

complex and unexpected patterns that would be dif-

ﬁcult or impossible to proﬁle using known data at-

tributes. Another advantage of using such an ap-

proach is that it is adaptive, i.e., the quality of the sys-

tem improves over time. The more available ratings,

the more accurate recommendations can be gener-

ated. Different from content-based recommendation,

CF can add a serendipitous factor into the recommen-

dation process. CF attracted much of attention in the

past decade, resulting in signiﬁcant progress and be-

ing adopted by some successful commercial systems,

including Amazon and Netﬂix.

3.2 Open Issues and Challenges

However, as the most commonly applied approach,

collaborative ﬁltering still provide less than accurate

rating prediction due to several open issues (Zanker

et al., 2011). In this study, we focus on the follow-

ing four problems, which we perceive to be the most

signiﬁcant ones.

1. Data Sparsity: Due to the fact that users typi-

cally rate or experience only a small fraction of

available items, the rating matrix for a RS with

millions of items is very sparse, which makes it

hard to ﬁnd users who have a similar rating be-

havior. Consequently, the quality of the generated

recommendations might suffer from this.

2. Cold Start: The term “cold start” in RS design

(Schein et al., 2002) basically describes two situ-

ations: (a) making recommendations to new users

that have not rated any item, and (b) dealing with

items that have not been rated or bought yet. RSs

are usually unable to make good quality recom-

mendations in such situations, since these newly

joined users cannot be linked with similar users,

so as the items. However, these are the users who

need good quality recommendations the most as

an incentive to continue using the system.

3. Rating Integrity: The integrity of ratings is an-

other problem in RS, due to lack of incentives

for a customer to rate products and the existence

of the malicious users. RSs are facing latent at-

tacks that aim to inﬂuence the functioning of the

system, including random attack, average attack,

bandwagon attack, segment attack love/hate at-

tack and so on (Mobasher et al., 2007; O’Mahony

et al., 2005; Hurley et al., 2007). Incentive design

for RS and robustness of the system under attack-

ers is currently an open issue that attracts more

ICISSP2015-DoctoralConsortium

and more research interest, especially when RSs

start to become more decentralized.

4. Scalability: Scalability of recommendation algo-

rithms with large and real-world datasets is also

a practical and important consideration in recom-

mender system design.

3.3 Trust-aware Social Recommender

System as a Possible Approach

Social network sites (SNS), like Facebook and Epin-

ions, provide as a complementary source of informa-

tion for recommendation. Thus the idea of integrat-

ing information from SNS into RS to improve the

performance of the recommendation algorithm (Guy

et al., 2009) has been commonly accepted and has at-

tracted an increasing amount of research (Palau et al.,

2004; He and Chu, 2010; Tekin et al., 2013). Many

social network-based recommender system designs

have been proposed and evaluated (Yang et al., 2014).

In this work, we are trying to develop a trust-aware

social recommender system, a social network-based

RS based on CF techniques, which takes the behav-

iors and preferences of neighbors in SNS into consid-

eration and interprets the social connections among

users as trust relationships. A trust network is set

up accordingly. The information extracted from SNS

about the users in the recommender system could

largely ﬁll the “white space” and tackle the problem

of cold start and data sparsity, as well as improve

the relevance of recommendations and user satisfac-

tion. We mainly focus on a static network struc-

ture while dynamic networks will be considered more

thoroughly in our further work.

4 STATE OF THE ART

4.1 Social Recommender Systems

Since RS users share more and more information

in associated social network sites (SNS), it makes

SNS an important source of information about users

and their preferences. SNS users have social pro-

ﬁle, which can be used to calculate high-dimensional

social and interest similarity between users. Mean-

while, due to the inter-connection of users in social

networks, the effect of social inﬂuence from neighbor

or community can be exploited in recommendation al-

gorithms. Several issues of RSs mentioned above, for

instance cold start, can be addressed with the help of

this additional source of information. A social recom-

mender system (SRS) can be built via integrating SNS

with the recommender system and using information

about users social proﬁles and/or relationships to sug-

gest items that might be of interest to them.

The concept of social recommender system has

been proposed for about a decade and a series of

approaches have been proposed concerning combin-

ing information from social network into RS for bet-

ter performance. However, the topic is still far from

well-development, waiting for better solutions. A col-

lective knowledge systems is proposed in (Gruber,

2008), in order to leverage collective intelligence in

semantic web. Bonhard et al. (Bonhard and Sasse,

2006) studied the factors that drive people’s decision-

making and advice-seeking through empirical studies,

and found out that the proﬁle similarity and rating

overlap of a recommender have a signiﬁcant impact

on a person’s decision. Carmagnola et al. (Carmag-

nola et al., 2009) proposed to evaluate the value of

a user’s interest in an item (”score”) using both the

strength of the relationship between the user and its

neighbors and the level of interest the item has for

each neighbor of the target user. The RS algorithm

designed in (He and Chu, 2010) integrates user’s own

preference, item’s general acceptance and inﬂuence

from friends. While the introduction of distant friends

improves the coverage of the recommendation algo-

rithm, it affects the accuracy of the approach.

4.2 Trust Evaluation in Social Network

When integrating a social network of directed links

to improve the performance of the recommender sys-

tem, these directed links can be interpreted as trust

relationships among users, which reﬂect the prefer-

ence (fondness), social closeness (subjective similar-

ity) and integrity of users in the network. These trust

relationships can be exploited to provide extra source

of information for recommendation hence improve

the performance of the RS. A lot of research has been

conducted in inferring and evaluating trust/distrust re-

lationships in SNS, as well as establishing trust net-

work based on SNS structure.

On calculating trust values, Richardson et al. pro-

posed to deploy a web of trust (Richardson et al.,

2003), where each user only maintains trusts of a

small number of other users (i.e. neighbors) and trust

values for all other users can be calculated via the

trust network. Zaihrayeu et al. presented an trust

inference infrastructure called IWtrust for calculat-

ing trust values for answers from the web (Zaihrayeu

et al., 2005). Golbeck proposed TidalTrust (Golbeck,

2005), which aggregates the weighted trust values be-

tween neighbors with direct trust. TidalTrust only

takes into account the shortest, strongest paths thus

Trust-awareSocialRecommenderSystemDesign

may loose some ratings from distant users in the net-

work. MoleTrust (Avesani et al., 2005) proposed by

Avesani et al. is similar to TidalTrust but considers

all raters up to a maximum-depth given as an input.

Weng et al. used trust for the cold start problem by au-

tomatic trust generation even for users who have rated

few items in common (Weng et al., 2006). Jøsang

et al. (Jøsang et al., 2006) treat trust and distrust as

two separate concepts and proposed three probabilis-

tic aggregation operators called consensus operators

for the fusion of dependent, independent and partially

dependent opinions respectively. However, these op-

erators assume that users have equal weights (equally

importance), and hence lack ﬂexibility. DuBois et al.

designed a probabilistic approach to infer the trust re-

lationship between users in social networks under ax-

ioms of inferred trust and applied it in network clus-

tering (DuBois et al., 2009b). In their following work

(DuBois et al., 2011), they further considered dis-

trust in the network and introduced a modiﬁed spring-

embedded algorithm for trust inference. Leskovec et

al. conducted research on signed edge (link) predic-

tion in online social networks using a machine learn-

ing framework (Leskovec et al., 2010). They showed

that the information about negative relationships is

useful in edge prediction.

4.3 Trust-aware Recommender System

Introducing a trust network into RS could improve not

only the coverage but also the accuracy of the recom-

mendation and the satisfaction of users. Also, since

malicious nodes can be recognized and excluded from

trust network, many attacks towards recommender

system can be effectively resisted. Currently RS are

only being used in low risk domains, due to lack of

transparency in RS. According to (Sinha and Swearin-

gen, 2001), users prefer more transparent systems,

and tend to rely more on recommendations from peo-

ple they trust than on those from anonymous people

“similar”to them. With the trust network, the RS can

be more transparent and easier to be accepted, open-

ing space for more application scenarios for RS.

In (Massa and Avesani, 2004), a trust-aware

method for recommender system is proposed, where

the collaborative ﬁltering process is informed by the

reputation of users computed via propagating trust.

The design was showed to increases the coverage

while not reducing the accuracy. Bedi et al. (Bedi

et al., 2007) proposed a trust-based recommender sys-

tem for the semantic web with the knowledge dis-

tributed over the network in the form of ontologies,

where a trust network is deployed to generate recom-

mendations. Andersen et al. (Andersen et al., 2008)

developed an natural set of ﬁve axioms, namely sym-

metry, positive response, independence of irrelevant

stuff, neighborhood consensus and transitivity de-

sired for designing trust-aware recommender systems,

and proposed a recommendation algorithm based on

random walks by weakening the axioms. In order

to tackle the problem of opinion ignorance and in-

consistency which may affect trust estimations, Vic-

tor et al. proposed a new framework called “Trust

Score”(Victor et al., 2011), where the trust relation is

a fuzzy mapping (Grifﬁths, 2006) from agent pairs to

trust vectors in [0, 1]

, containing trust degree and dis-

trust degree. Additional dimensionality of trust value

introduced in (Victor et al., 2011) can differentiate

partial trust, partial distrust, partial ignorance and par-

tial inconsistency. In (DuBois et al., 2009a) the author

used an probabilistic trust inference algorithm devel-

oped in (DuBois et al., 2009b) to set up trust metrics

and conduct trust-based clustering on users to further

improve recommendation accuracy.

Some approaches also tried to combine classic

collaborative ﬁltering with trust-based recommenda-

tion algorithms. O’Donovan et al.(O’Donovan and

Smyth, 2005) focused on trust-based adaptations of

collaborative ﬁltering, which they called trust-based

ﬁltering, where only the most trustworthy neighbors

participate in the recommendation process. Two trust-

aware methods at proﬁle-level and proﬁle-item-level

respectively are introduced to improve standard col-

laborative ﬁltering methods and showed that trust in-

formation can help increase recommendation accu-

racy. This algorithm does not involve trust propa-

gation or aggregation thus has the disadvantage of

sparser rating matrix. Ma et al. demonstrated that a

factor analysis method (Ma et al., 2008), which fuses

the user-item matrix with the users’ social trust net-

works, generates better recommendations than the tra-

ditional collaborative ﬁltering algorithms. This ap-

proach, however, does not reﬂect the real world rec-

ommendation process. Later in (Ma et al., 2009), a

user’s ﬁnal rating decision was interpreted as the bal-

ance between this user’s own taste and her trusted

users’ favors, and an ensemble probabilistic matrix

factorization method is proposed to implement the

idea. Jamali et al. (Jamali and Ester, 2009) developed

a random walk model combining the trust-based and

the collaborative ﬁltering approach for recommenda-

tion. The random walk model allows to deﬁne and to

measure the conﬁdence of a recommendation. Wei et

al. proposed a multi-collaborative ﬁltering trust net-

work algorithm (MCFTN) for recommendation un-

der Web 2.0 circumstance, with the assumption that

all the necessary information is available (Wei et al.,

2013). In their approach, different sources of infor-

ICISSP2015-DoctoralConsortium

mation are combined in a collaborative ﬁltering fash-

ion. However, the trust propagation does not consider

aggregation of different opinions.

5 METHODOLOGY

In order to tackle the issues mentioned in Sec. 3.2,

we propose a trust-aware recommender system which

integrates information from directed social networks.

Trust is an umbrella term for a wide range of

meanings. In the research area of RS, trust can be

interpreted in several different ways:

• Trust can be deﬁned as a measure of conﬁdence

that the user’s rating reﬂects her real opinion

(i.e. integrity), which is evaluated to discover and

avoid attacks on the system. This is similar to

the deﬁnition of trust in control and operation re-

search (Theodorakopoulos and Baras, 2006; Jiang

and Baras, 2009).

• Trust can also be used to describe the quality of

the recommendations made by the system.

• When integrating SNS into RS, trust is directed

relationship between users and is a compound of

integrity, preference and social closeness. In such

circumstance, trust among users forms a network

called trust network (Victor et al., 2011).

In this literature, we use the third deﬁnition of trust in

our trust-aware recommender system.

5.1 Trust Model

The main aim in setting up trust networks is to al-

low agents to form trust opinions on unknown agents

or sources by asking for a trust opinion from ac-

quainted agents. While trust is increasingly involved,

the use and modeling of distrust remains relatively un-

explored. Although recent research (Golbeck, 2008;

Victor et al., 2011) show an emerging interest in mod-

eling the notion of distrust, models that take into ac-

count both trust and distrust are still scarce. Most ap-

proaches completely ignore distrust, or consider trust

and distrust as opposite ends of the same continuous

scale. However, there is a growing body of opinion

that distrust cannot be seen as the equivalent of lack

of trust (Gans et al., 2001; Guha et al., 2004).

Referring to previous work (Massa and Bhat-

tacharjee, 2004; Guha et al., 2004; Golbeck, 2005;

Massa and Avesani, 2007; Golbeck, 2008; Walter

et al., 2008; Victor et al., 2011), we deﬁne the concept

of trust network used in our trust-aware recommender

system as follows.

Deﬁnition. Trust Network: The directed trust net-

work T (V, E) is established via the social graph asso-

ciated with the recommender system, where V is the

set of nodes (i.e. users) with |V | = N, E is the set of

directed edges (i.e. trust links). ∀ directed edge e

i j

, v

) ∈ E, v

, v

∈ V , is a directed trust link from

node v

towards v

, with value (weight) w

i j

∈ [−1, 1]

indicating the extent of trust that node v

put on v

These links are not necessarily symmetric. A weight

of 1 means “totally agree”or “like”, while weight

of −1 means “totally disagree”or “dislike”. N

i j

∈ E} is the neighbor set of node v

As is in classic collaborative ﬁltering, there exist a

set of items O = {o

, o

, ··· , o

}. The rating of user

on item o

is r

, which is the element in ith column

and kth row of the rating matrix R. The ratings are

integers and the rating range is set to be [−5, 5], both

conﬁgurable to different application scenarios.

Trust networks are typically challenged by two

important problems that inﬂuence trust opinions.

Firstly, in large networks it is likely that many agents

do not know each other, hence there is an abundance

of ignorance. Secondly, because of the lack of a cen-

tral authority, different agents might provide differ-

ent and even contradictory information, thus incon-

sistency may occur. In our trust model, we propose

to apply trust propagation to eliminate the ignorance

on users due to no direct connections. Regarding the

second problem, we design a distributed trust aggre-

gation rule that can handle both trust and distrust.

Deﬁnition. Trust Propagation: In our trust-aware

system, trust between two nodes of no direct con-

nections can be estimated using trust values of edges

along the path between the two nodes. A maximum

path length λ and trust threshold τ are set up to save

computation resource and avoid inﬁnite loop. If the

path length exceeds λ or the trust value decreases be-

low τ along the path, then there’s no trust (or distrust)

relation between the two nodes.

Figure 1: Trust propagation.

Fig. 1 illustrates how the trust propagation works

in our system. In the example, v

can reach indirect

trust (w

) about v

via its direct trust (w

) about v

and v

’s direct trust about v

(i.e. w

). If the max-

imum length λ ≥ 2, and the threshold τ ≤ w

· w

then v

’s indirect trust about v

exists and can be cal-

culated as Eq. 2 (otherwise w

doesn’t exist).

= w

· w

(2)

Trust-awareSocialRecommenderSystemDesign

When there are multiple paths between two nodes,

the indirect trust value calculated along different paths

are combined using trust aggregation. There are two

major ways to conduct trust aggregation, namely First

Aggregate Then Propagate (FATP) and First Propa-

gate Then Aggregate (FPTA). We apply the ﬁrst ap-

proach (i.e. FATP) in our system.

Deﬁnition. Trust Aggregation: Indirect trust values

towards node v

calculated from different paths can

be combined in a recursive way: for nodes that have

direct trust values about v

, the aggregated trust val-

ues are their direct trust values; for each node v

along

the paths who has no direct trust value, she learns her

neighbors’ trust values about v

and combine them ac-

cording to her trust towards her neighbors:

∑

∈N

i j

≥σ

i j

· w

∑

∈N

i j

≥σ

i j

(3)

The opinion of neighbor v

∈ N

of v

will not be

take into consideration if v

has low trust about v

(i.e.

i j

< σ the threshold).

Figure 2: Trust aggregation.

Fig. 2 illustrates how the trust aggregation is con-

ducted in our system. If w

, w

and w

are all

greater than the threshold σ, they will be used for cal-

culating the trust of v

about v

using Eq. 4

· w

+ w

· w

+ w

· w

+ w

(4)

5.2 Trust-aware Recommendation

Algorithms

Based on our trust model, we propose several can-

didate trust-aware recommendation algorithms which

takes both trust and distrust into consideration.

5.2.1 Distrust Filtering

For RS of adequate rating information, we propose to

integrate trust network via Distrust-based Filtering.

This approach is intuitive and easy to apply to col-

laborative ﬁltering RS. The idea is that with the trust

network, we could ﬁlter out users of low trust even

they are “similar” to the target user.

There are two possible approaches:

• If the target user’s trust value about a node is lower

than the pre-conﬁgured threshold, the node’s rat-

ing will not be considered in rating prediction.

• When using neighborhood-based CF method, the

calculated similarity will be weighted by the trust

about the node.

When applying the ﬁrst approach, S(k; i) should

be updated as {v

∈ S(k; i), w

i j

≥ η}. When using

the second approach, Eq. 1 should be modiﬁed as:

ˆr

= b

∑

∈S(k;i),w

i j

≥0

i j

· (r

− b

)

∑

∈S(k;i),w

i j

≥0

i j

(5)

In this way, the security of the RS can be im-

proved. However, the coverage of the RS has no im-

provement and the problem of data sparsity and cold

start are not addressed via this approach.

5.2.2 Trust-weighted Recommendation

Similar to previous work (Massa and Avesani, 2004;

Golbeck, 2005; Massa and Avesani, 2007; Golbeck,

2008), this method uses trust relationships for rec-

ommendations. When introducing distrust (negative

opinion), the equation for rating prediction is modi-

ﬁed accordingly. Here we propose a 2-step approach.

1. In the ﬁrst step, only trusted neighbors are con-

sidered in rating prediction.

2. Then the opinions of distrusted nodes are mirrored

about the rating reached in the ﬁrst step. Then cal-

culate the weighted average rating with absolute

value of trust as weights.

Based on our trust propagation and aggregation

rule, this recommendation algorithm has a greedy im-

plementation. For node v

, if she doesn’t have direct

rating about item o

, then in step 1, her preliminary

rating ¯r

about item o

can be calculated using Eq. 6:

¯r

∑

∈N

i j

≥0

i j

· r

∑

∈N

i j

≥0

i j

(6)

With the preliminary rating ¯r

, a rating r

about

from a distrusted node v

(with negative trust

value) is adjusted as ˜r

according to Eq. 7.

˜r

= min{max{2 ¯r

− r

, r

min

}, r

max

} (7)

where the min and max bounds are used to avoid ex-

ceeding the rating range. With the adjusted rating of

ICISSP2015-DoctoralConsortium

distrusted nodes, in the second step, the predicted rat-

ing about r

can be calculated with Eq. 8.

ˆr

∑

∈N

i j

≥0

i j

· r

∑

∈N

| · ˜r

∑

∈N

i j

≥0

i j

∑

∈N

≥0

(8)

In order to avoid inﬁnite loops and save computa-

tion resources, the length of the trust propagation path

is upper-bounded by K.

Note that in this algorithm, only information of

neighbors in the trust network is needed to reach the

prediction on ratings, thus the algorithm has the ad-

vantage of good scalability in large scale systems.

The number of trusted neighbors used in calculation,

similar to user-based CF, is conﬁgurable and can be

tuned to reach better performance.

5.2.3 Recommendation with Random Trust

Propagation

Even through coverage of the RS can be improved

via trust propagation, the accuracy of recommenda-

tion algorithm may be affected. In order to balance

the coverage and accuracy, the major concerns in RS,

we follow the idea of using ratings of similar items

with probabilistic switch between trust propagation

and approximate rating discussed in (Jamali and Es-

ter, 2009), and proposed our recommendation algo-

rithm with probabilistic trust propagation based on

the trust-aware recommendation algorithm discussed

in Sec. 5.2.2.

At each node v

, we don’t always predict its rat-

ing about an item o

via aggregating opinions from its

neighbors in N

. Instead, there exists probability θ

ik,t

that the rating ˆr

is calculated with the v

’s ratings

about items that are similar to o

. Here the similarity

metric can be any form, e.g. cosine similarity used in

item-based collaborative ﬁltering.

As the probability to choose between trust prop-

agation and rating prediction with similar items, θ

ik,t

is related to node v

, item o

and the position of v

along the trust propagation path (i.e. the step t). If v

chooses to use her ratings on similar items to predict

ˆr

, then

ˆr

∑

∈I(k;i)

· r

∑

∈I(k;i)

(9)

where I(k; i) is the set of items rated by v

which are

similar to o

Apart from the upper-bound K for trust propaga-

tion, another condition is introduced to terminate the

propagation process:

(1 − θ

ik,t

) · max{w

i j

, j ∈ N

} ≥ Γ (10)

where Γ is the threshold for the condition and is tun-

able. This condition indicates that if v

has rated items

very similar to o

(θ

ik,t

very large), or the neighbors

of v

are not trustworthy, then the trust propagation

process will end. The performance of the system is

determined by paths of high trust and items of high

similarity with the target item, both related to Γ.

The switch probability θ

ik,t

can also be interpreted

as a weight term, in which case the equation for rating

prediction is:

ˆr

=(1 − θ

ik,t

)

∑

∈N

i j

≥0

i j

· r

∑

∈N

| · ˜r

∑

∈N

i j

≥0

i j

∑

∈N

≥0

+ θ

ik,t

∑

∈I(k;i)

· r

∑

∈I(k;i)

(11)

Similar to the trust-aware recommendation al-

gorithm described in Eq. 8, this probabilistic trust

propagation model can also be implemented in de-

centralized way and only local information is used to

reach prediction, which makes it highly scalable in

large scale systems.

5.2.4 Trust-aware Combinatorial CF

In this method, both similarity and trust values are

normalized. A proportion α is calculated based on the

total weight of similar nodes and trusted neighbors:

∑

∈S(k;i),w

≥η

∑

∈S(k;i),w

≥η

∑

∈T

i j

(12)

where S(k;i) is the neighbor set of v

about item o

is the set of nodes used for trust-aware recommen-

dation for v

, and η is the threshold used to exclude

users who is distrusted by v

The combinatorial CF is a mixture of user-based

and trust-aware collaborative ﬁltering. The propor-

tion α

is the weight of result reached via user-based

CF, and the estimated rating can be expressed as:

ˆr

=(1 − α

) · ˆr

trust

+ α

· ˆr

sim

=(1 − α

)

∑

j∈N

i j

≥0

i j

∑

n∈N

| ˜r

∑

j∈N

i j

≥0

i j

∑

n∈N

+ α

∑

l∈S(k;i),w

≥η

− b

)

∑

l∈S(k;i),w

≥η

)

(13)

where ˆr

sim

is the prediction of user-based CF (Her-

locker et al., 1999; Koren, 2008), and ˆr

trust

is the

result of trust-aware CF as mention in Sec. 5.2.2.

5.2.5 Trust-aware Composite CF

With normalized similarity and trust values, we can

apply the way EnsembleTrustCF (Victor et al., 2011)

used to combine ratings of similar users and trusted

neighbors:

Trust-awareSocialRecommenderSystemDesign

ˆr

∑

∈N

i j

≥0

i j

− b

) +

∑

∈N

|( ˜r

− b

)

∑

∈S(k;i)

i j

∑

∈N

i j

≥0

i j

∑

∈N

≥0

∑

∈S(k;i)

i j

− b

)

∑

∈S(k;i)

i j

∑

∈N

i j

≥0

i j

∑

∈N

≥0

+ b

(14)

5.2.6 Trust Value Update

’s trust value about node v

can be updated via com-

paring the predicted ratings ˆr

, real ratings r

and r

as shown in Fig. 3. When ˆr

locates between r

and

, it means j’s opinion makes negative contributions

to recommendation, thus the trust value w

i j

will de-

crease (w

i j

= w

i j

−δ). When r

lies in between r

and

ˆr

, the weight w

i j

should be increased (w

i j

= w

i j

+ δ),

since j makes positive contribution in this case.

Figure 3: Update w

i j

based on real and predicted ratings.

With such trust update scheme, the trust values are

adjusted according to users’ (nodes) behaviors and the

prediction results using trust-aware recommendation

algorithms be more accurate and the performance of

the system can be further improved. This trust dynam-

ics is distributed and restricted only to neighbors , in-

stead of requiring users to maintain global knowledge.

Such distributed trust update scheme makes the sys-

tem efﬁcient in memory and computation resources.

6 EXPECTED OUTCOME

Current recommender systems are confronting prob-

lems like data sparsity, cold start and adversary at-

tacks. One promising approach is to establish a trust

network from the social networks associated with the

RS and conduct rating prediction with neighbors’ rat-

ings in the trust network. Such an RS design is called

trust-aware social recommender system.

In this work, we propose several trust-aware rec-

ommendation algorithms, in order to increase knowl-

edge base for better recommendations, by introduc-

ing trust signals from SNS. Specially, we introduce a

novel way to handle negative trust (distrust). We also

consider integrating the novel recommendation algo-

rithms with classic collaborative ﬁltering approach.

By introducing information from social network,

as shown in Sec. 5, the knowledge base of the rec-

ommender system can be enlarged and the issue of

data sparsity can be addressed. Thus we expect that

the performance of the RS can be improved. Espe-

cially for the case of cold start, when the system has

little information about the user’s preference, it’s hard

for classic collaborate ﬁltering methods to predict her

ratings over items. However, with trust relationships

extracted from social connections, the system is ex-

pected to be able to predict her preferences and offer

proper suggestions on items she needs.

Among the ﬁve approaches in trust-aware RS de-

sign that we propose in this work, two are solely

based on trust values and ratings of neighbors in trust

network and can be implemented in a decentralized

way, which makes the two approaches highly scal-

able. Here in our model, the trust propagation ends

within a limited number of steps, which is inspired

by the fact that social inﬂuence is shallow (Ugander

et al., 2011). We predict that this setting still offers

relatively precise trust estimation, meanwhile bring-

ing a substantial save on computation resource and

boost on algorithm efﬁciency. Besides, since trust

value update doesn’t require global information, it en-

ables the prompt update on trust values, which could

further improve the accuracy of the algorithms.

Since the trust relationship between users are ex-

ploited in the algorithm, it’s much harder for the ad-

versaries to attack or compromise the system with

techniques used on current recommender systems.

Thus we expect our system to be more secure.

We will conduct experiments on large scale

datasets to test the performance of the trust-aware

recommendation algorithms that are proposed in this

work and compare them with other state-of-art meth-

ods, in order to verify our design and adjust the model

to further improve the performance.

So far, we assume that the social network associ-

ated with the recommender system does not change

over time, which in reality may not be valid. Typi-

cally, social networks evolve over time. Thus in our

further work, we plan to investigate the inﬂuence of

network dynamics to the system, and introduce prob-

abilistic model to describe the system behaviors.

Trust is currently deﬁned as a 1-dimensional

value. In the future we plan to extend it to a multi-

dimensional vector, with each element representing

the trust relation in that dimension (e.g. category).

Meanwhile, we will discuss the robustness of the dif-

ferent RS designs under adversary attacks. We are

also interested in expanding the horizon of recom-

mender systems and making them available in more

application scenarios.

ICISSP2015-DoctoralConsortium

REFERENCES

Adomavicius, G. and Tuzhilin, A. (2005). Toward the

next generation of recommender systems: A survey of

the state-of-the-art and possible extensions. Knowl-

edge and Data Engineering, IEEE Transactions on,

17(6):734–749.

Andersen, R., Borgs, C., Chayes, J., Feige, U., Flaxman,

A., Kalai, A., Mirrokni, V., and Tennenholtz, M.

(2008). Trust-based recommendation systems: an ax-

iomatic approach. In Proceedings of the 17th interna-

tional conference on World Wide Web, pages 199–208.

ACM.

Avesani, P., Massa, P., and Tiella, R. (2005). A trust-

enhanced recommender system application: Moleski-

ing. In Proceedings of the 2005 ACM symposium on

Applied computing, pages 1589–1593. ACM.

Bedi, P., Kaur, H., and Marwaha, S. (2007). Trust based

recommender system for semantic web. In IJCAI, vol-

ume 7, pages 2677–2682.

Bonhard, P. and Sasse, M. (2006). knowing me, know-

ing youusing proﬁles and social networking to im-

prove recommender systems. BT Technology Journal,

24(3):84–98.

Burke, R. (2007). Hybrid web recommender systems. In

The adaptive web, pages 377–408. Springer.

Carmagnola, F., Vernero, F., and Grillo, P. (2009). Sonars:

A social networks-based algorithm for social recom-

mender systems. In User Modeling, Adaptation, and

Personalization, pages 223–234. Springer.

DuBois, T., Golbeck, J., Kleint, J., and Srinivasan, A.

(2009a). Improving recommendation accuracy by

clustering social networks with trust. Recommender

Systems & the Social Web, 532:1–8.

DuBois, T., Golbeck, J., and Srinivasan, A. (2009b). Rig-

orous probabilistic trust-inference with applications to

clustering. In Web Intelligence and Intelligent Agent

Technologies, 2009. WI-IAT’09. IEEE/WIC/ACM In-

ternational Joint Conferences on, volume 1, pages

655–658. IET.

DuBois, T., Golbeck, J., and Srinivasan, A. (2011). Pre-

dicting trust and distrust in social networks. In Pri-

vacy, security, risk and trust (passat), 2011 ieee third

international conference on and 2011 ieee third inter-

national conference on social computing (socialcom),

pages 418–424. IEEE.

Gans, G., Jarke, M., Kethers, S., and Lakemeyer, G. (2001).

Modeling the impact of trust and distrust in agent net-

works. In Proc. of AOIS01, pages 45–58.

Golbeck, J. (2008). Computing with social trust. Springer.

Golbeck, J. A. (2005). Computing and applying trust in

web-based social networks.

Grifﬁths, N. (2006). A fuzzy approach to reasoning with

trust, distrust and insufﬁcient trust. In Cooperative

Information Agents X, pages 360–374. Springer.

Gruber, T. (2008). Collective knowledge systems: Where

the social web meets the semantic web. Web seman-

tics: science, services and agents on the World Wide

Web, 6(1):4–13.

Guha, R., Kumar, R., Raghavan, P., and Tomkins, A.

(2004). Propagation of trust and distrust. In Proceed-

ings of the 13th international conference on World

Wide Web, pages 403–412. ACM.

Guy, I., Zwerdling, N., Carmel, D., Ronen, I., Uziel, E.,

Yogev, S., and Ofek-Koifman, S. (2009). Personalized

recommendation of social software items based on so-

cial relations. In Proceedings of the third ACM confer-

ence on Recommender systems, pages 53–60. ACM.

He, J. and Chu, W. W. (2010). A social network-based rec-

ommender system (SNRS). Springer.

Herlocker, J. L., Konstan, J. A., Borchers, A., and Riedl,

J. (1999). An algorithmic framework for perform-

ing collaborative ﬁltering. In Proceedings of the

22nd annual international ACM SIGIR conference on

Research and development in information retrieval,

pages 230–237. ACM.

Hurley, N. J., O’Mahony, M. P., and Silvestre, G. C. (2007).

Attacking recommender systems: A cost-beneﬁt anal-

ysis. Intelligent Systems, IEEE, 22(3):64–68.

Jamali, M. and Ester, M. (2009). Trustwalker: a random

walk model for combining trust-based and item-based

recommendation. In Proceedings of the 15th ACM

SIGKDD international conference on Knowledge dis-

covery and data mining, pages 397–406. ACM.

Jiang, T. and Baras, J. S. (2009). Graph algebraic interpre-

tation of trust establishment in autonomic networks.

Preprint Wiley Journal of Networks.

Jøsang, A., Marsh, S., and Pope, S. (2006). Exploring dif-

ferent types of trust propagation. In Trust manage-

ment, pages 179–192. Springer.

Koren, Y. (2008). Factorization meets the neighborhood:

a multifaceted collaborative ﬁltering model. In Pro-

ceedings of the 14th ACM SIGKDD international con-

ference on Knowledge discovery and data mining,

pages 426–434. ACM.

Leskovec, J., Huttenlocher, D., and Kleinberg, J. (2010).

Predicting positive and negative links in online social

networks. In Proceedings of the 19th international

conference on World wide web, pages 641–650. ACM.

Ma, H., King, I., and Lyu, M. R. (2009). Learning to rec-

ommend with social trust ensemble. In Proceedings

of the 32nd international ACM SIGIR conference on

Research and development in information retrieval,

pages 203–210. ACM.

Ma, H., Yang, H., Lyu, M. R., and King, I. (2008). Sorec:

social recommendation using probabilistic matrix fac-

torization. In Proceedings of the 17th ACM conference

on Information and knowledge management, pages

931–940. ACM.

Massa, P. and Avesani, P. (2004). Trust-aware collaborative

ﬁltering for recommender systems. In On the Move

to Meaningful Internet Systems 2004: CoopIS, DOA,

and ODBASE, pages 492–508. Springer.

Massa, P. and Avesani, P. (2007). Trust-aware recommender

systems. In Proceedings of the 2007 ACM conference

on Recommender systems, pages 17–24. ACM.

Massa, P. and Bhattacharjee, B. (2004). Using trust in

recommender systems: an experimental analysis. In

Trust Management, pages 221–235. Springer.

Trust-awareSocialRecommenderSystemDesign

Mobasher, B., Burke, R., Bhaumik, R., and Williams,

C. (2007). Toward trustworthy recommender sys-

tems: An analysis of attack models and algorithm ro-

bustness. ACM Transactions on Internet Technology

(TOIT), 7(4):23.

O’Donovan, J. and Smyth, B. (2005). Trust in recommender

systems. In Proceedings of the 10th international con-

ference on Intelligent user interfaces, pages 167–174.

ACM.

O’Mahony, M. P., Hurley, N. J., and Silvestre, G. C. (2005).

Recommender systems: Attack types and strategies.

In AAAI, pages 334–339.

Palau, J., Montaner, M., L

opez, B., and De La Rosa, J. L.

(2004). Collaboration analysis in recommender sys-

tems using social networks. In Cooperative Informa-

tion Agents VIII, pages 137–151. Springer.

Richardson, M., Agrawal, R., and Domingos, P. (2003).

Trust management for the semantic web. In The Se-

mantic Web-ISWC 2003, pages 351–368. Springer.

Schein, A. I., Popescul, A., Ungar, L. H., and Pennock,

D. M. (2002). Methods and metrics for cold-start rec-

ommendations. In Proceedings of the 25th annual in-

ternational ACM SIGIR conference on Research and

development in information retrieval, pages 253–260.

ACM.

Sinha, R. R. and Swearingen, K. (2001). Comparing recom-

mendations made by online systems and friends. In

DELOS workshop: personalisation and recommender

systems in digital libraries, volume 106.

Tekin, C., Zhang, S., and van der Schaar, M. (2013). Dis-

tributed online learning in social recommender sys-

tems. arXiv preprint arXiv:1309.6707.

Theodorakopoulos, G. and Baras, J. S. (2006). On trust

models and trust evaluation metrics for ad hoc net-

works. Selected Areas in Communications, IEEE

Journal on, 24(2):318–328.

Ugander, J., Karrer, B., Backstrom, L., and Marlow, C.

(2011). The anatomy of the facebook social graph.

arXiv preprint arXiv:1111.4503.

Victor, P., Cornelis, C., and De Cock, M. (2011). Trust net-

works for recommender systems, volume 4. Springer.

Walter, F. E., Battiston, S., and Schweitzer, F. (2008). A

model of a trust-based recommendation system on a

social network. Autonomous Agents and Multi-Agent

Systems, 16(1):57–74.

Wei, C., Khoury, R., and Fong, S. (2013). Web 2.0 rec-

ommendation service by multi-collaborative ﬁltering

trust network algorithm. Information Systems Fron-

tiers, 15(4):533–551.

Weng, J., Miao, C., and Goh, A. (2006). Improving collab-

orative ﬁltering with trust-based metrics. In Proceed-

ings of the 2006 ACM symposium on Applied comput-

ing, pages 1860–1864. ACM.

Yang, X., Guo, Y., Liu, Y., and Steck, H. (2014). A survey

of collaborative ﬁltering based social recommender

systems. Computer Communications, 41:1–10.

Zaihrayeu, I., Da Silva, P. P., and McGuinness, D. L. (2005).

Iwtrust: Improving user trust in answers from the web.

In Trust Management, pages 384–392. Springer.

Zanker, M., Felfernig, A., and Friedrich, G. (2011). Recom-

mender systems: an introduction. Cambridge Univer-

sity Press.

ICISSP2015-DoctoralConsortium