Trust based recommender system using ant colony for trust computation


Trust based recommender system using ant colony for trust computation

Punam Bedi, Ravish Sharma ⇑
Department of Computer Science, Faculty of Mathematical Sciences, Opposite Daulat Ram College, University of Delhi, Delhi 110007, India

a r t i c l e i n f o

Keywords:
Collaborative Filtering
Trust
Recommender system
Ant colony
Pheromone updating

a b s t r a c t

Collaborative Filtering (CF) technique has proven to be promising for implementing large scale recom-
mender systems but its success depends mainly on locating similar neighbors. Due to data sparsity of
the user–item rating matrix, the process of finding similar neighbors does not often succeed. In addition
to this, it also suffers from the new user (cold start) problem as finding possible neighborhood and giving
recommendations to user who has not rated any item or rated very few items is difficult. In this paper,
our proposed Trust based Ant Recommender System (TARS) produces valuable recommendations by
incorporating a notion of dynamic trust between users and selecting a small and best neighborhood
based on biological metaphor of ant colonies. Along with the predicted ratings, displaying additional
information for explanation of recommendations regarding the strength and level of connectedness in
trust graph from where recommendations are generated, items and number of neighbors involved in pre-
dicting ratings can help active user make better decisions. Also, new users can highly benefit from pher-
omone updating strategy known from ant algorithms as positive feedback in the form of aggregated
dynamic trust pheromone defines ‘‘popularity’’ of a user as recommender over a period of time. The per-
formance of TARS is evaluated using two datasets of different sparsity levels viz. Jester dataset and
MovieLens dataset (available online) and compared with traditional Collaborative Filtering based
approach for generating recommendations.

� 2011 Elsevier Ltd. All rights reserved.

1. Introduction

The complexity of the recommendation problem is due to its vast
space of possibilities. Recommender systems are important tools
that overcome the information overload by sifting through the large
set of data and recommending information relevant to the user. Rec-
ommendation techniques have a number of possible classifications
including content based, collaborative, knowledge-based, demo-
graphic, and utility based (Burke, 2002; Resnick & Varian, 1997;
Schafer, Konstan, & Reidl, 1999; Terveen & Hill, 2001). Collaborative
Filtering, Content based approach and Hybrid methods are the pre-
valent three approaches to developing recommender systems. In
Collaborative Filtering (CF) approach, opinions from users in the
form of ratings on various items are collected. The recommendations
produced are based only on the opinions of users similar to the active
user. Active user refers to recommendation seeker for whom recom-
mendations are generated. Content based approach suggests items
that are similar to the ones the active user has shown a preference
for in the past rather than on the preferences of other users. Mostly
CF technique is combined with one or more recommendation

techniques i.e. content based, knowledge-based, demographic, or
utility based and is called Hybrid approach.

CF takes its roots from something humans have been doing for
centuries i.e. sharing opinions with others and brings together the
opinions of large interconnected communities on the web (Schafer,
Frankowski, Herlocker, & Sen, 2007). Collaborative Filtering based
recommenders work best for a user who fits into a niche with
neighbors of similar taste. This approach doesnot need a represen-
tation of the items in terms of features but is based only on the
judgment of the user community. Because of its simplicity in both
theory and implementation, CF can be applied to virtually any kind
of item viz. papers, news, web sites, movies, songs, books, jokes,
locations of holidays, stocks etc. The three-step conceptual model
of the operation of the Collaborative Filtering process is as follows:

(i) Construct rating matrix using user–item data.
(ii) Locate people (neighbors) with similar profiles (similar

preferences).
(iii) Unify neighbors’ ratings to form recommendations.

There are several limitations and challenges to CF based recom-
mender systems due to traditional emphasis on user similarity.
One of the main limitations is that it provides recommendations
based solely on the opinion of the users whose preferences best
matches the taste of active user. Approaches incorporating trust

0957-4174/$ - see front matter � 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.eswa.2011.07.124

⇑ Corresponding author. Address: B-408, Flex Apartments, Plot No. C-58/22,
Sector 62, Noida 201301, U.P., India. Tel.: +91 09313907606; fax: +91 11 27662553.

E-mail addresses: pbedi@cs.du.ac.in (P. Bedi), sharma_ravish123@rediffmail.com
(R. Sharma).

Expert Systems with Applications 39 (2012) 1183–1190

Contents lists available at ScienceDirect

Expert Systems with Applications

j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e s w a

http://dx.doi.org/10.1016/j.eswa.2011.07.124
mailto:pbedi@cs.du.ac.in
mailto:sharma_ravish123@rediffmail.com
http://dx.doi.org/10.1016/j.eswa.2011.07.124
http://www.sciencedirect.com/science/journal/09574174
http://www.elsevier.com/locate/eswa


models into recommender systems are gaining momentum, syn-
thesizing recommendations based upon opinions from trusted
peers rather than most similar ones (Ziegler & Nilcolas, 2007). In
online recommendation generation process, a number of users
form a virtual social network where trust relationship is the key
of users’ human relations and trust relationship of mutual interde-
pendence forms a so-called Web of trust (Wei, 2007). The three
properties assigned to trust as discussed in literature (Ries, Kanga-
sharju, & Muhlhauser, 2006) are as follows:

(i) Trust is subjective and therefore asymmetric.
(ii) Trust is context dependent.

(iii) Trust is dynamic which means that it can increase with posi-
tive experience and decrease with negative experience or
over time without any experience.

The dynamic property of trust can be viewed as trust intensity
among the users. For example, the trust intensity between recom-
mender and active user may increase or decrease depending on
recommendations generated by the recommender. Trust intensity
being time-based information, can be analyzed using pheromone
updating strategy known from ant algorithms. Ant algorithms are
based on emulation of behavior of real ants and build better solu-
tions by communicating through artificial pheromone. In many ant
species, ants tend to lay a substance called pheromone while walk-
ing from their nests to food source and vice versa. Ants do not com-
municate directly with each other, but they follow pheromone
trails laid by other ants. Ants are attracted by pheromones coming
from fellow type (members of same species) ants and repulsed by
pheromone of non-fellow type ants. As the time passes, paths that
are marked by stronger amount of pheromone are chosen with
higher probability than those that have weaker amount of phero-
mone deposit (Blum, 2005; Bonabeau, Dorigo, & Theraulaz, 1999;
Dorigo, Maniezzo, & Colorni, 1996; Dorigo & Stutzle, 2004). One
of the basic ingredients of the ant algorithms is a randomized gree-
dy construction heuristic in which each option is selected with a
probability proportional to its perceived quality. As the algorithm
progresses, the probabilities are altered to encourage the selection
of elements that have featured good solutions. Ant algorithms
repeatedly construct new solutions from scratch throughout the
duration of the search. Moreover, this collaborative behavior be-
tween fellow type ants has an analogy with the collaborative world
as people mostly collect opinions from their like-minded friends,
neighbors etc. Work by Bedi, Sharma, and Kaur (2009) and Sharma,
Singh, Makkar, Kaur, and Bedi (2007) combine artificial pheromone
information with similarity measure for choosing best matching
clusters providing active user with good set of alternative recom-
mendations. In addition to the taste of active user, clusters that
are marked by stronger amount of pheromone have the higher
probability of being chosen than those that have weaker amount
of pheromone deposit. Also, one of the fundamental challenges
for recommender systems is to improve the quality of the pre-
dicted ratings. In such a scenario, pheromone updating strategy
of ants guides the search to a better recommendation. The positive
feedback in the form of pheromone deposition results in achieving
an emergent, unified behavior for the recommender system as a
whole, and produces a robust system capable of finding improved
quality recommendations. Still, there is a need for producing valu-
able recommendations due to data sparsity of input-rating matrix
and solving new user problem. These problems are intrinsic in the
process of finding similar neighbors. In our proposed approach
TARS (Trust based Ant Recommender System), sparseness in user
similarity due to data sparsity of input matrix is reduced while
creating directed trust graph for each user. Weight on the edge rep-
resents strength of connectedness i.e. trust intensity between the
two recommendation partners (recommender and active user) at

time t. Due to its dynamic property, it is called Dynamic Trust Pher-
omone (DTP) and is updated using pheromone updating strategy
known from ant algorithms. As the time passes, recommendations
are produced by continuously updating dynamic trust between
users and selecting a small and best neighborhood using ant colony
metaphor. Also, new users can highly benefit from pheromone
updating strategy known from ant algorithms.

The rest of the paper is organized as follows: Section 2 reviews
the related research work and presents the motivation for our
work. Our proposed system TARS with a detailed description of
each step is presented in Section 3. Section 4 explains the experi-
mental methodology employed and the two datasets used in our
experiments. Section 5 concludes the paper.

2. Related work

In this section, review of literature related to recommender sys-
tems and motivation for our work is presented.

Many collaborative systems have been developed in the acade-
mia and the industry. Using stereotypes, the Grundy system was
developed to build individual user models which were used to rec-
ommend relevant books to each user (Rich, 1979). Later on, the
Tapestry system relied on each user to identify like-minded users
manually (Goldberg, Nicholas, Oki, & Terry, 1992). GroupLens
(Konstan et al., 1997; Resnick, Lakovou, Sushak, Bergstrom, & Riedl,
1994) in the domain of Usenet newsgroup articles, Bellcore’s Video
Recommender (Hill, Stead, Rosenstein, & Furnas, 1995) in the
domain of movies and Ringo (Shardanand & Maes, 1995) in the do-
main of music and musical artists were the first systems to use Col-
laborative Filtering algorithms to automate predictions. Other
examples of collaborative recommender systems include the book
recommendation system from Amazon.com, the PHOAKS system
that helps people find relevant information on the WWW (Terveen,
Hill, Amento, McDonald, & Creter, 1997), and the Jester system that
recommends jokes (Goldberg, Roeder, Gupta, & Perkins, 2001). Dif-
ferent item-based recommendation algorithms are analyzed by
Sarwar, Karypis, Konstan, and Riedl (2001).

Researchers have also applied Swarm Intelligence techniques to
recommender systems. Ujjin and Bentley (2003) have described
Particle Swarm Optimization (PSO) recommender system in which
PSO algorithm has been employed to learn personal preferences of
users and provide tailored suggestions. A system called CASIS has
been developed combining case based reasoning approach with a
metaphor from colonies of social insects, namely the honey bee
dance. This combination has been used in the retrieval step of
the recommendation cycle (Lorenzi, Santos, & Bazzan, 2005).
Web-based system user interface hybrid recommendation method
is presented in (Sobecki, 2007) where ant colony metaphor is used
for selecting the most optimal path in the user interface graph.

In addition to traditional emphasis on user similarity, trustwor-
thiness of users has been an important consideration by research-
ers. They have suggested that the advantage of combining users’
trust network and user–item rating matrix solves CF sparseness
problem to some extent. Two computational models of trust
namely profile-level trust and item level trust have been developed
and incorporated into standard Collaborative Filtering frameworks
(O’Donovan and Smith, 2005). Work by Massa and Bhattacharjee
(2004) and Massa and Avesani (2009, 2007, 2004) focus on using
the explicit trust as input along with user–item rating matrix to
predict ratings. They have analyzed explicit trust from the popular
Internet web site epinions.com and shown that by exploiting the
web of trust, it is possible to propagate trust and infer an additional
weight for other users. New users can also benefit from trust prop-
agation as long as the users provide at least one trusted friend.

Although the explicit trust based recommender system models
have high rating prediction coverage and high rating prediction

1184 P. Bedi, R. Sharma / Expert Systems with Applications 39 (2012) 1183–1190



https://isiarticles.com/article/7769