HYBRIDISING COLLABORATIVE FILTERING
AND TRUST-AWARE RECOMMENDER SYSTEMS
Charif Haydar
1
, Anne Boyer
2
and Azim Roussanaly
2
1
Womup, University Nancy 2, Tours, Nancy, France
2
Loria Laboratory, University Nancy 2, Nancy, France
Keywords:
Recommender Systems, Trust, Reputation, Users Similarity.
Abstract:
Recommender systems (RS) aim to predict items that users would appreciate, over a list of items. In evaluation
of recommender systems, two issues can be defined: accuracy of prediction which implies the satisfaction of
the user, and coverage which implies the percentage of satisfied users. Collaborative filtering (CF) is the master
approach in this domain, but still has some weaknesses especially about coverage. Trust-aware approach is
today another promising approach in RS within social environments, whose prediction exceeds the quality of
(CF). In this paper we propose several strategies to hybridize both approaches in order to improve prediction
accuracy and coverage.
1 INTRODUCTION
Recommender systems (RS) (Resnick and Varian,
1997) aim to recommend users some items they
would appreciate, over a list of items. Collaborative
filtering (CF) (Resnick et al., 1994) is the most popu-
lar approach of recommender systems. The main idea
behind CF is to recommend to a user, called current
user, the items appreciated by users who are similar
to him/her in the term of preferences. These similar
users are called neighbors in this context.
CF suffers from many weaknesses, such as: cold
start (Maltz and Ehrlich, 1995), data sparsity (Shin
et al., 2008), fragility to malicious attacks (Mobasher
et al., 2007; Burke et al., 2005) and recommendation
acceptability by users (Herlocker et al., 2000).
In social systems, users can not only express their
opinions about items, but also about other users,
whereby they rate their credibility and trustworthi-
ness. The literature has proposed to replace user
similarity by trust relationships, which resulted in
trust-aware recommenders (Massa and Bhattacharjee,
2004; Golbeck and Hendler, 2006). Trust-aware ap-
proaches have the advantages of alleviating the prece-
dent weaknesses of CF recommenders, without bring-
ing the recommendation accuracy down (Massa and
Bhattacharjee, 2004). What these systems really of-
fer to their users is the possibility to choose their list
of neighbors themselves, these neighbors are called
friends in this context. Choosing friends manually
comes out to more accurate lists. Nevertheless, a size-
able percentage of users are still out of this social
range. They still rate items but not other users. Even
though trust-aware approaches surpass user similarity
in recommendation quality they are unable to recom-
mend items to those users, while users similarity ap-
proaches are still able to. This fact leads us to think
that hybridising both approaches will allow the RS to
recommend to a larger percentage of users.
Hybrid recommenders (Burke, 2007), which hy-
bridise several recommendation approaches to gen-
erate recommendation, were widely proposed as so-
lutions to the case when different approaches suffer
from contradictory weaknesses. This allows to profit
of their advantages together. To the best of our knowl-
edge, opinion similarity and trust similarity have al-
ways been considered as independent notions. The
main contribution of this paper is to exploit both no-
tions altogether. We propose to hybridise them in one
recommender system in the aim of yielding it more
profitable by a larger set of users, ensuring a minimal
recommendation accuracy.
The outline of the paper is organized as follows: in
section 2 we present the user-based collaborative fil-
tering recommenders, and trust-aware recommenders.
In section 3 we address to the hybridization strategies
to extract a set of them applicable our case. Last sec-
tion is dedicated to discussing the experiments and re-
sults.
695
Haydar C., Boyer A. and Roussanaly A..
HYBRIDISING COLLABORATIVE FILTERING AND TRUST-AWARE RECOMMENDER SYSTEMS.
DOI: 10.5220/0003937406950700
In Proceedings of the 8th International Conference on Web Information Systems and Technologies (WEBIST-2012), pages 695-700
ISBN: 978-989-8565-08-2
Copyright
c
2012 SCITEPRESS (Science and Technology Publications, Lda.)
2 GENERAL FRAMEWORK
2.1 Recommender Systems
Recommender systems utilize a wide range of tech-
niques in order to recommendpertinent items to users,
the skeleton part of the recommender system is the
prediction function. This function estimates how
much a given user will like a given item. A predic-
tion function might exploit information about other
users (user based) (Resnick et al., 1994) or informa-
tion about other items (item based) (Basu et al., 1998)
or both of them (hybrid) (Burke, 2007) to generate the
recommendations.
In this paper we are interested in user based rec-
ommenders, and more specifically in collaborativefil-
tering and trust-aware recommender systems.
2.2 Collaborative Filtering
Recommenders
CF is the most popular recommendation approach.
The prediction function is based on the similarity of
users’ ratings.These ratings are stored in a m× n ma-
trix called rating matrix, where m is the number of
users and n is the number of items. An element v
u
a
i
of
this matrix is the rating user u
a
to the item i. Users’
similarity is computed by using this matrix. Many
metrics in the litterature satisfy this definition of sim-
ilarity in the CF recommenders (Resnick et al., 1994;
Breese et al., 1998).
In this paper, we will consider the Pearson cor-
relation coefficient (Resnick et al., 1994) as a sim-
ilarity metrics, which varies between the values 1
(completely opposite users) and +1 (completely sim-
ilar users). Our choice is found on the wide popularity
and the efficiency of this metrics.
It is defined in the equation 1:
f
simil
(u
a
, u
j
) =
iI
a
T
I
j
(v
(u
a
,i)
v
u
a
)(v
(u
j
,i)
v
u
j
)
q
iI
a
T
I
j
(v
(u
a
,i)
v
u
a
)
2
iI
a
T
I
j
(v
(u
j
,i)
v
u
j
)
2
(1)
Where:
u
a
, u
j
are two users.
v
(u
a
,i)
: is the rating of user u
a
to the item i.
v
u
a
: is the average rate of the user u
a
.
In order to predict how much the current user u
a
will rate an item r the system exploits the rating of
similar users to u
a
(equation 2) from the group of
users who rated r (U
r
).
p(u
a
, r) =
v
u
a
+
u
j
U
r
f
simil
(u
a
, u
j
) × (v
(u
j
,i)
v
u
j
)
card(U
r
)
(2)
Where:
U
r
: the set of users who have rated r.
card(U
r
): is the number of users in U
r
.
CF is incapable to generate accurate recommen-
dations to new users who have not rated items yet.
Paradoxically, the system needs to attract this kind of
users (Massa and Bhattacharjee, 2004). This context
is called the cold start problem. Data sparsity prob-
lem is observed in datasets containing large archives
of items, where even an active user can’t rate 1% of
the total number of items. By consequence, the ma-
jority of the rating matrix values are empty. In sparse
matrix, the probability that users rate the same items
becomes smaller. As a result, similarity between users
becomes rare and the accuracy of the prediction be-
comes very low.
2.3 Trust Aware Recommenders
The study of trust as a computational phenomenon
started in the last decade. Many models were pro-
posed (Abdul-Rahman, 2004; Kruknow, 2006; Mui,
2002). We are interrested in the qualitative definition
of trust (called also local-trust), where personal bias
is taken into account, and trust is represented as user
to user relationship (Ziegler and Lausen, 2004b).
A correlation between trust and users similarity
was found in (Ziegler and Lausen, 2004a) and (Lee
and Brusilovsky, 2009). Replacing user similarity by
trust relationships is the object of several ”trust-aware
RS” (Golbeck and Hendler,2006; Massa and Avesani,
2004). In a social network, such as eopinion
1
and
filmtrust
2
, when a user A expresses that he trusts an-
other user B RS considers that B is eligible to recom-
mend items to A.
Trust-aware systems apply the same prediction
function as in CF, with only one difference, which is
replacing similarity values by trust values. The trust
values can be given explicitly by the current user, or
computed by a trust propagation algorithm.
Using trust in RSs allows to alleviate several of
CF’s weaknesses. A cold start user has only to express
at least one trust relation to start to receive recommen-
dation, this is, anyway, less costly than rating dozen of
items. The propagation algorithms allow to increase
the connectivity of users. This can reduces the im-
pact of the data sparsity. While users choose them-
selves their friends this allows to easily identify ma-
1
http://www.eopinion.com
2
http://trust.mindswap.org/FilmTrust/
WEBIST2012-8thInternationalConferenceonWebInformationSystemsandTechnologies
696
licious attackers and to isolate them to protect other
users. Finally trust aware recommendation is clearer
and easier to explain to the user, which improves the
acceptability.
Many models were presented to model trust and
it’s propagation (Massa and Bhattacharjee, 2004;
Golbeck, 2005; Ziegler and Lausen, 2004b; Kuter and
Golbeck, 2010). All those models consider trust prop-
agation problem as a graph traversal problem. The
difference between those models is about their strate-
gies in traversing the graph and the choice of path be-
tween the source and destination nodes.
In most models, trust is expressed either as real or
integer value within a given range, whereas it is a bi-
nary value in our studied case. That is why we choose
for our experiments the model MoleTrust (Massa and
Bhattacharjee, 2004). In MoleTrust, each user has a
domain of trust where he adds his trustees, that is to
say, user can either trusts fully other user or he does
not trust him at all. The only initializing parameter is
the maximal propagation distance d.
If user A added user B to his domain, and B added
C, then the trust of A in C is given by the equation:
Tr(A,C) =
(d n+ 1)
d
Where n is the distance between A and C (n = 2 as
there two steps between them; first step from A to B,
and the second from B to C).
3 HYBRIDISATION
In (Burke, 2007) the author identifies seven strategies
to hybridise multiple recommendation approaches, he
argues that there is no reason why recommenders
from the same type could not be hybridized. Some
of these strategies require that at least one of the
hybridised approaches to be an item-based recom-
mender which is beyond the scope of this paper, other
strategies require that the approaches are applied in
separated recommenders, while in some strategies the
hybridisation is done in the prediction function level.
In this paper we are interested in the last type only.
We present here the chosen strategies:
Weighted: The score of both similarity and
trust are combined numerically according to
predefined weights:
score(u
a
, u
j
) =
α× simil(u
a
, u
j
) + (1 α) × trust(u
a
, u
j
)
(3)
Mixed: Each approach produces her list of rec-
ommendation independently, the lists then are
mixed and send to the user as one list. As for our
case:
score(u
a
, u
j
) = max(simil, trust)
Probabilistic: This strategy privileges neighbors
who are both similar and trustee over those who
are either only similar or only trustee, with respect
to the similarity and the trust values. The formula
here is:
score = 1 (1 simil)(1 trust)
Switching: The system selects a single recom-
mendation approach, and switch to the second one
only when the quality of recommendation of the
first one is not satisfactory.
score(u
a
, u
j
) =
simil simil 6= null
trust simil = null
Cascade: The idea here is to create a strictly hi-
erarchical hybrid, in which a weak recommender
can not overturn the decisions taken by a stronger
one, but can merely refine them. In order to be se-
lected by the prediction function the user must be
similar to and trusted by the active user. In other
words; the first approach chooses trustee users,
the second approach chooses similar users within
the trustees set.
The score that we apply in this case is:
score(u
a
, u
j
) =
trust simil > 0
0 otherwise
Notice that the last two strategies are sensitive to
the order of hybridised approaches, so we shall test
each of them twice with alterning the approaches’ or-
der.
In order to have the similarity and trust value in the
same range, all similarity values in the experiments
are normalized.
4 EXPERIMENTS
4.1 DataSet
We use the eopinion.com dataset. eopinion.com is
a consumers opinion website where users can rate
items in a range of 1 to 5, and write reviews about
them. Users can also express their trust towards re-
viewers whose reviews seem to be interesting to them.
Eopinion dataset contains 49,290 users who rated a
total of 139,738 items. The total number of ratings
is 664,824. The dataset contains also 487,182 binary
HYBRIDISINGCOLLABORATIVEFILTERINGANDTRUST-AWARERECOMMENDERSYSTEMS
697
trust ratings. It is important also to mention that 3,470
users have neither rated an item nor trusted a user,
these users are eliminated from our statistics. Thus
the final number of users is 45,820 users.
4.2 Methodology of Evaluation
In (Massa, 2006), authors showed on this corpus how
does replacing similarity metrics by trust-aware met-
rics improve both accuracy and coverage. The im-
provementof coverage was limited because of the fact
that some users are active in rating items but not in
rating reviewers. 11,858 users have not trusted any-
body in the site (25.8% of the total number of users).
Those users have made 75,109 ratings, on average of
6.3 ratings by user. This high average value means
that recommendations can be generated to this cate-
gory of users by a similarity based approach, and not
by a trust-aware apraoch. On the other hand, 5,655
users have not rated any item in the site (12.3% of the
total number of users). The average of trust relation
by user in this set is 4.07 which is not negligible, those
users suffer from the same problem with the similar-
ity approaches while trust based approach can predict
item to them. We are convinced that each of similar-
ity and trust approaches is suitable to a particular set
of users. This is why we think that hybridising these
approaches can satisfy a larger set of users, while pro-
viding accurate recommendations.
Our objective is to find the suitable hybridisation
formula, that leads to improve the satisfaction of the
system. The satisfaction in this context is represented
by both accuracy and coverage.
We divide the corpus in two parts randomly, 80%
for training and 20% for evaluation (a classical ratio in
the literature). We respected this ratio by user, so ev-
ery user has 80% of his ratings in the training corpus
and 20% in the evaluation corpus, this is important
when we want to measure satisfaction by user.
In each experiment, a rating matrix is formed from
the training corpus, while the ratings of the evaluation
corpus are considered as empty values. The recom-
mender must predict those values and complete the
matrix with them. the resulting matrix is compared to
the original full rating matrix (formed from the entire
corpus ratings). The evaluation of the recommender
depends on how many values could it predict (cover-
age), and how much the predicted values are close to
the real ones (accuracy).
To measure accuracy, we use the mean absolute
error metrics (MAE) (Herlocker et al., 2004), which
is a widely used predictive accuracy metrics. MAE
measures the average absolute deviation between the
predicted values of ratings and the real values sup-
plied by the user.
MAE focuses on ratings but not on users (Massa
and Avesani, 2004). User mean absolute error
(UMAE) (Massa and Avesani, 2004) is the version of
MAE which consider the accuracy by user. It consists
in computing the MAE to the predictions of every
user. And then computing the average of all MAEs.
The aim of RS is to predict appreciable items to
user. A prediction function that succeeds in predict-
ing low and middle ratings of the user but fail in pre-
dicting high ones can not generate recommendations.
High MAE (Baltrunas, 2007) is the version of MAE
dedicated to evaluate the ability of the system to rec-
ommend and not to predict. This version of MAE
takes into account only high ratings (usually 4 and 5
for systems of rating range between 1 and 5).
It is also vital to know the percentage of ratings
and users the system can cover. We employ three
forms of coverage metrics, compatible with the three
forms of MAE. These metrics are:
Coverage of prediction: the number of predicted
ratings divided by the total number of ratings in
the evaluation corpus.
Coverage of users: the number of users who re-
ceived predictions divided by the total number of
users.
Coverage of high-ratings: the number of predicted
high ratings divided by the total number of high
ratings in the evaluation corpus.
4.3 Results and Discussion
The weighted hybridisation is the only proposed strat-
egy to have a parameter (α). To adjust this parameter,
we associated it with 5 different values between 0.1
and 0.9 with a step of 0.2. Table 1 illustrates the MAE
changes according to α values. It is obvious that the
change is slight. For what concern the coverage it
is stable for any value of α, which is normal while
changing α affects the value of the prediction but not
the ability to predict. Nevertheless, we will choose
the value of α = 0.3 in our upcoming comparisons,
because it has the best MAE score (MAE = 0.821).
Table 1: α and MAE for weighted hybridisation strategy.
α 0.1 0.3 0.5 0.7 0.9
MAE 0.8219 0,821 0,8212 0,823 0,8294
All results are illustrated in table 2, to facilitate the
analysis we group each couple of metrics in a figure
while discussing them.
Figure 1 illustrates the MAE and the coverage of
prediction metrics for all tested strategies. Almost all
WEBIST2012-8thInternationalConferenceonWebInformationSystemsandTechnologies
698
hybridisation strategies improve the prediction cover-
age by approximately 7% compared to the trust-aware
approach, and about 15% compared to similarity ap-
proach. The only exception is the cascade strategy,
although its accuracy is not very far from that of other
strategies, it has a very low coverage. This is because
cascade is severe in eliminating neighbors.
This result shows how the hybridsystem uses each
approach to predict ratings inpredictable by the other
approach, which allows to cover much ratings than
any of the two approaches.
Figure 1: MAE and coverage.
It is obvious now that hybridisation ensures more
predictions, the next question is what kind of predic-
tions? and what is the impact of this improvement on
the recommendation?
As we argued above, HMAE is a metrics that
replies to this question. Figure 2 shows that hybridi-
sation enhence the ability to predict good items by
about 5% compared to MoleTrust and 12% compared
to Pearson correlation similarity. Once again we ex-
clude the cascade.
This result shows that hybrid system can rec-
ommend more items than trust-aware and similarity
based recommenders, with ensuring the approximate
accuracy to that of trust-aware approach, and farther
better than similarity based approach.
Figure 2: HMAE and high-ratings coverage.
Another important question is: Does this augmen-
tation in prediction coverage concern a limited num-
ber of users?
Figure 3 shows the values of UMAE by strate-
gies. Hybridisation strategies cover 10% of users
more compared to MoleTrust and 15% more com-
pared to Pearson similarity. This ratio is not negli-
gible and shows that satisfaction is shared by a con-
siderable propotion of users, so hybrid systems are
capable to predict to a large variety of users
Notice that in cascade, ordering trust before sim-
ilarity allowed us to gain 0.08 in UMAE. In the first
strategy we use the trust score of similar users, while
already similars are rare, not a lot of trustees are avail-
able. In the second, we use the similarity of trustee
users, while trustees are usually numerous, this give
more chance to find similar among them.
Figure 3: UMAE and users’ coverage.
Finally, our results show that hybridisation in-
creases coverage in RS, with keeping the accuracy
in a reasonable range. This improvement includes
ratings predictions, recommendations and users. It
also adapts with larger variety of users behavior. For
what concern the choice of hybridisation strategy, it
is likely to employ those who tend to aggregate in-
formation of many approaches, and avoid severity in
eliminating information.
5 CONCLUSIONS AND FUTURE
WORK
In this paper we showed that even though trust-aware
recommenders can usually improve the accuracy and
the coverage of prediction of CF recommenders over
usual similarity metrics, it is still unable to recom-
mend to ceratain categories of users. Hybridising
these approaches is a promising strategy to improve
the coverage without significant decrease in accuracy.
Our results where proved using one corpus, which
mean that it can be specific to it. We are convinced
that validation must be done on other corpora when
available. The nature of current corpus restricted
our choice of trust metrics. We hope that upcoming
tests to be done on datasets with numeric trust values,
which allow to test other trust metrics.
Finally, similarity metrics allows the detection of
non similar users. Current trust metrics do not treat
HYBRIDISINGCOLLABORATIVEFILTERINGANDTRUST-AWARERECOMMENDERSYSTEMS
699
Table 2: The three forms of MAE and coverage.
Strategy MAE coverage HMAE High coverage UMAE User coverage
Pearson correlation 0.84 61.15% 0.6364 44.44% 0.8227 47.46%
MoleTrust 0.8165 69.28% 0.6185 51.12% 0.8079 52.21%
Cascade (Simil,Trust) 0.8315 53.19% 0.6263 38.57% 0.8905 36.78%
Cascade (Trust,Simil) 0.8358 53.44% 0.6339 38.94% 0.8126 36.97%
Mixed 0.8208 76.34% 0.6218 56.43% 0.8124 62.15%
Probabilistic 0.8206 76.31% 0.6212 56.44% 0.8124 62.07%
Switch (Trust,Simil) 0.8217 76.38% 0.622 56.19% 0.8161 61.96%
Switch (Simil,Trust) 0.8220 76.38% 0.623 56.44% 0.8148 62.06%
Weighted (α = 0.3) 0.8210 76.38% 0.6214 56.42% 0.8124 62.22%
the distrust issue. We would like to extend our work
to integrate this aspect which we find important.
REFERENCES
Abdul-Rahman, A. (2004). A Framework for Decentralised
Trust Reasoning. phD thesis. PhD thesis, Computer
Sceince, University College London.
Baltrunas, R. (2007). Dynamic item weighting and selec-
tion for collaborative filtering. In Web mining 2.0
Workshop.
Basu, C., Hirsh, H., and Cohen, W. (1998). Recommenda-
tion as classification: using social and content-based
information in recommendation. In Proceedings of
the fifteenth national/tenth conference on Artificial in-
telligence/Innovative applications of artificial intelli-
gence.
Breese, J. S., Heckerman, D., and Kadie, C. (1998). Empir-
ical analysis of predictive algorithm for collaborative
filtering. In Proceedings of the 14 th Conference on
Uncertainty in Artificial Intelligence, pages 43–52.
Burke, R. (2007). Hybrid web recommender systems. In
Brusilovsky, P., Kobsa, A., and Nejdl, W., editors,
The adaptive web, pages 377–408. Springer-Verlag,
Berlin, Heidelberg.
Burke, R., Mobasher, B., Zabicki, R., and Bhaumik, R.
(2005). Identifying attack models for secure recom-
mendation. In in Beyond Personalization: A Workshop
on the Next Generation of Recommender Systems.
Golbeck, J. (2005). Personalizing applications through inte-
gration of inferred trust values in semantic web-based
social networks.
Golbeck, J. and Hendler, J. (2006). FilmTrust: movie rec-
ommendations using trust in web-based social net-
works. In Consumer Communications and Network-
ing Conference, 2006. CCNC 2006. 3rd IEEE.
Herlocker, J. L., Konstan, J. A., and Riedl, J. (2000).
Explaining collaborative ltering recommendations.
Proceedings of the 2000 ACM conference on Com-
puter supported cooperative work CSCW 00.
Herlocker, J. L., Konstan, J. A., Terveen, L. G., John, and
Riedl, T. (2004). Evaluating collaborative filtering
recommender systems. ACM Transactions on Infor-
mation Systems, 22:5–53.
Kruknow, K. (2006). Towards of trust for the global Ubiq-
uitous Computer. PhD thisis. PhD thesis, University
of Aartus.
Kuter, U. and Golbeck, J. (2010). Using probabilistic con-
fidence models for trust inference in web-based social
networks. ACM Trans. Internet Technol.
Lee, D. H. and Brusilovsky, P. (2009). Does trust influence
information similarity?
Maltz, D. and Ehrlich, K. (1995). Pointing the way: active
collaborative filtering. In Proceedings of the SIGCHI
conference on Human factors in computing systems.
Massa, A. (2006). Trust-aware bootstrapping of recom-
mender systems. In ECAI Workshop on Recommender
Systems.
Massa, P. and Avesani, P. (2004). Trust-aware collaborative
filtering for recommender systems. In In Proc. of Fed-
erated Int. Conference On The Move to Meaningful
Internet: CoopIS, DOA, ODBASE, pages 492–508.
Massa, P. and Bhattacharjee, B. (2004). Using trust in rec-
ommender systems: An experimental analysis. In
iTrust’04, pages 221–235.
Mobasher, B., Burke, R., Bhaumik, R., and Williams, C.
(2007). Toward trustworthy recommender systems:
An analysis of attack models and algorithm robust-
ness. ACM Transactions on Internet Technology.
Mui, L. (2002). Computional Models of Trust and Repu-
tation: Agents, Evolutionary Games,and Social Net-
works. PhD thesis. PhD thesis, Massechusetts Insti-
tute of Technology.
Resnick, P., Iacovou, N., Sushak, M., Bergstrom, P., and
Riedl, J. (1994). Grouplens: An open architecture for
collaborative filtering of netnews. In 1994 ACM Con-
ference on Computer Supported Collaborative Work
Conference.
Resnick, P. and Varian, H. R. (1997). Recommender sys-
tems. Commun. ACM, 40(3):56–58.
Shin, H., Kim, N. Y., Kim, E. Y., and Lee, M. (2008).
Behaviors-based user profiling and classification-
based content rating for personalized digital tv. In
Consumer Electronics, 2008. ICCE 2008. Digest of
Technical Papers. International Conference on.
Ziegler, C.-N. and Lausen, G. (2004a). Analyzing cor-
relation between trust and user similarity in online
communities. In Proceedings of Second International
Conference on Trust Management, pages 251–265.
Springer-Verlag.
Ziegler, C.-N. and Lausen, G. (2004b). Spreading activation
models for trust propagation. In Proceedings of the
2004 IEEE International Conference on e-Technology,
e-Commerce and e-Service (EEE’04).
WEBIST2012-8thInternationalConferenceonWebInformationSystemsandTechnologies
700