Towards a Generic Architecture for Recommenders Benchmarking
Mohamed Ramzi Haddad
1
, Hajer Baazaoui
1
, Djemel Ziou
2
and Henda Ben Ghezala
1
1
Riadi-Gdl Laboratory, École Nationale des Sciences de l’Informatique, Université de la Manouba, Manouba 2010, Tunisia
2
Centre de Recherche MoIVRe, Département d’Informatique, Faculté des Sciences, Université de Sherbrooke
2500 Boul. Université, Sherbrooke, J1K 2R1, Canada
Keywords:
Hybrid Recommender Systems, User Modeling, Consumption Behaviors Prediction, Benchmarking.
Abstract:
With current growth of internet sales and content consumption, more research efforts are focusing on devel-
oping recommendation and personalization algorithms as a solution for the choice overload problem. In this
paper, we first enumerate several state-of-the-art recommendation algorithms in order to highlight their main
ideas and methodologies. Then, we propose a generic architecture for recommender systems benchmarking.
Using the proposed architecture, we implement and evaluate several variants of existing recommendation al-
gorithms and compare their results to our unified recommendation model. The experiments are conducted on
a real world dataset in order to assess the genericity of our recommendation model and its quality. At the end,
we conclude with some ideas for further development and research.
1 INTRODUCTION
As e-commerce and content delivery businesses are
increasingly popular on the web, it became neces-
sary to address the information overload problem
and provide an intelligent and personalized access to
the available goods and content. Several social and
consumption psychology studies such as (Schwartz,
2005), (Jacoby et al., 1974) and (Iyengar and Lepper,
2000) confirmed the existence of an overchoice prob-
lem (also known as choice overload) and stressed the
need to reduce customers’ choices to facilitate the de-
cision making.
Scientific and industrial communities proposed
several tools to address this challenge such as person-
alization and recommendation systems. The goal of a
recommender system is to propose a set of interesting
items for a user based on heuristics or on acquired
knowledge. Suggestions of books on Amazon and
movies on Netflix are real world examples of recom-
mender systems. Each recommendation algorithm is
based on a set of assumptions about consumption be-
haviors in order to predict individuals’ interests and
future purchases. Therefore, system designers may
rely on hybridization of several recommendation al-
gorithms in order to take into consideration all the
facts that drive consumers’ purchase decisions and in-
terests. In this context, we believe that it may be more
relevant to propose a new class of multi-faceted rec-
ommendation algorithms, capable of adapting their
recommendations based on the domain-driven con-
sumption behaviors.
In this paper, we present a new generic contextual
recommendation model encompassing the main ideas
and the hypothesis of state of the art algorithms. Such
model would be able to describe and predict more pre-
cisely consumers interests and purchases by using all
the available data in the targeted field. Besides, we
develop a generic recommendation architecture that
is used to implement, study and assess the quality
state-of-the-art algorithms by integrating all the nec-
essary components for content recommendation. The
proposed recommendation model is also implemented
within the architecture in order to evaluate its perfor-
mances.
This paper is organized as follows. Next, in Sect.
2, we present the major state of the art of recom-
mendation approaches and describe briefely our pre-
vious works on statistical context-aware recommen-
dation models. Afterwards, in Sect. 3 we present a
generic architecture for recommendation algorithms
benchmarking. Finally, in Sect. 4, the experimenta-
tion context and the obtained results are presented and
discussed. The paper is concluded by summarizing
our proposition and presenting some future research
perspectives.
435
Ramzi Haddad M., Baazaoui H., Ziou D. and Ben Ghezala H..
Towards a Generic Architecture for Recommenders Benchmarking.
DOI: 10.5220/0005216904350442
In Proceedings of the International Conference on Agents and Artificial Intelligence (ICAART-2015), pages 435-442
ISBN: 978-989-758-074-1
Copyright
c
2015 SCITEPRESS (Science and Technology Publications, Lda.)
2 RECOMMENDATION
ALGORITHMS OVERVIEW
AND MOTIVATIONS
Recommender systems are a class of personalization
systems whose main objective is to predict users’ in-
terests towards the available informational content in
the application domain. To achieve this goal, several
approaches and methodologies were proposed in the
literature.
Collaborative filtering (CF) is widely adopted in e-
commerce and is based on the assumption that users
who, in the past, had the same attitudes towards
the same items would eventually agree in the future
(Goldberg et al., 1992). In user-based CF, as de-
tailed in (Herlocker et al., 1999), current user’s miss-
ing ratings are predicted by aggregating the ratings
of a neighborhood of similar users. Due to a lack
of scalability of this approach, item-to-item CF was
proposed. Hereby, instead of matching similar users,
the algorithm matches a user’s rated items to simi-
lar items. Item-based CF approaches, as proposed in
(Linden et al., 2003) and (Deshpande and Karypis,
2004), are based on the assumption that users pre-
fer items that are rated similarly, or correlated to the
items they already know and like. In practice, item-
based CF leads to faster systems and delivers a better
recommendation quality (cf. (Sarwar et al., 2001)).
Several researches consented to modeling item’s
features in the recommendation data model which led
to content-based filtering (CBF) approaches. CBF fo-
cuses on eliciting users’ preferences towards items’
features in order to use them to score unknown items.
Thus, in (Mooney and Roy, 2000), (Balabanovi
´
c
and Shoham, 1997) and (Pazzani, 1999), the recom-
mended items are the ones that have the most inter-
esting features regarding the user past interactions.
However, users’ preferences can be augmented by
their likeminded neighbors’ preferences in a collab-
orative filtering manner (Berkovsky et al., 2008).
Demographic filtering (DF) techniques adopt a
generalized user and item representations. In fact,
users are described by a set of demographic attributes
for a better handling of users similarity. Moreover,
items may be described by their features as in CBF
approaches. DF generalizes CF and CBF and may
then reuse their recommendation generation method-
ologies (Krulwich, 1997; Pazzani, 1999).
Several researches on recommender systems
pointed out the importance of the context in indi-
viduals’ choices and perceived relevancy. This led
to a new class of context-aware recommendation ap-
proaches that are able to identify cases where the con-
text (mainly time and location) implies some common
consumption behaviors and may then respond with
more accurate recommendations (Woerndl and Groh,
2007; Boutemedjet and Ziou, 2008).
Finally, several recent researches focus on hybrid
recommendation approaches. Recommender systems
hybridization relies on the assumption that the ag-
gregation of several recommendation techniques im-
proves their efficiency and helps overcoming their
shortcomings. In (Burke, 2007), the author argues
hybrid recommender systems and defines several hy-
bridization strategies.
In previous works, we proposed a new unified
contextual recommendation approach based on a sta-
tistical model of consumers behaviors (Haddad et al.,
2012). The model predicts users purchases and inter-
ests based on a set of factors issued from consumers
psychology researches without being restricted to any
of the underlying assumptions of existing approaches.
To achieve this goal using the proposed model, we
compute the probability of observing a user u assign-
ing a rating e to an item x in the context q and then
recommend the top N ones with the highest expected
rating (Haddad et al., 2012). The proposed model uni-
fies the main ideas of existing recommendation ap-
proaches and is able to dynamically select the most
appropriate inference technique based on probabili-
ties learned from existing data.
Based on the study of consumer psychology re-
searches and existing recommendation approaches,
we believe that system designers would benefit from a
generic architecture integrating all the necessary com-
ponent for content recommendation. Our work is mo-
tivated by the fact that such architecture would make
it possible to benchmark existing recommendation al-
gorithms and develop more generic ones that encom-
pass all the relevant ideas, methodologies and prac-
tices of the application domain. In this context, the
originality of this work consists on the development
of a generic architecture for content recommendation
and on the benchmarking of existing algorithms in or-
der to validate our unified contextual recommendation
model.
3 A GENERIC ARCHITECTURE
FOR RECOMMENDERS
BENCHMARKING
In this work, we propose a generic architecture for
recommender systems benchmarking in order to eval-
uate existing algorithms and compare their results to
our unified recommendation model. In this section,
first the main components of our architecture are de-
ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence
436
Figure 1: Generic recommendation architecture.
tailed (c.f. figure 1). Then, the implemented algo-
rithms and their different variants are enumerated.
3.1 Architecture Components
Figure 1, represents the components of the proposed
architecture which are the following :
1. Data repository : regroups the raw data on which
the recommendation process is based (e.g. users’
demographics, products features, ratings, etc. . . ).
2. Clustering algorithms: this component regroups
a set of clustering algorithms such as k-means
(Kaufman and Rousseeuw, 1990) and c-means
(Bezdek, 1981) for users and items clustering.
3. Recommendation models: this component acts as
a repository for recommendation models that are
generated by model-based approaches and used to
infer recommendations.
4. Recommendation engines: are responsible for
recommendations generation either from raw data
or from previously learned model. In our case,
recommendation engines are used to infer users’
ratings in order to recommend the items that are
the most likely to interest the active user.
5. Data collection processes: is responsible for col-
lecting users’ navigation data which may indicate
their interests (e.g. ratings, purchases, consulta-
tions, etc. . . ). Such data needs to be captured and
stored in order to be used as an input for the rec-
ommendation algorithms.
6. Quality measures: several quality measures may
be used in order to evaluate and compare the rec-
ommendation algorithms. In this work, the Mean
Absolute Error (MAE) and the Root Mean Square
Error (RMSE) measures were implemented to
evaluate the quality of the predicted ratings.
3.2 Implemented Algorithms
In this work, four recommendation approaches were
implemented and compared to our model. Hereby, the
objective is to predict the rating r
ac
that an active user
u
a
would assign to a candidate item x
c
. The imple-
mented algorithms are the following:
TowardsaGenericArchitectureforRecommendersBenchmarking
437
3.2.1 User-centered Collaborative Filtering
(UC-CF)
UC-CF as detailed in (Herlocker et al., 1999) and
(Resnick et al., 1994) is based the assumption that
users with similar preferences will rate products sim-
ilarly. Thus, the implemented variants of UC-CF un-
dergo the following stages:
• Compute USim(u
a
, u
i
) = USim(
→
r
a
and
→
r
i
), the
similarities between the active user u
a
and all the
other users u
i
based on their common ratings
→
r
a
and
→
r
i
assigned to the same items.
• Select S
u
a
, the set of the k most similar users to
the active user u
a
.
• Estimate ¯r
ac
using a rating estimator such as the
mean or the weighted mean (c.f. equations 2 and
3) on the set of ratings assigned by S
u
a
to x
c
.
3.2.2 Item-centered Collaborative Filtering
(IC-CF)
IC-CF approaches such as proposed in (Sarwar et al.,
2001) and (Deshpande and Karypis, 2004) are based
on the assumption that users prefer items that are cor-
related to the items they already know and like. Thus,
all the implemented IC-CF algorithms persue the fol-
lowing generic pattern:
• Compute ISim(x
c
, x
i
) = ISim(
→
r
c
,
→
r
i
), the similari-
ties between the candidate item x
c
and all the other
items x
i
based on the ratings
→
r
c
and
→
r
i
they were
assigned by the same users.
• Select S
x
c
, the set of the k most similar items to
the candidate item x
c
. S
x
c
is also referred to as the
neighborhood of x
c
.
• Estimate ¯r
ac
by aggregating the ratings assigned
by u
a
to the items in S
x
c
using a rating estima-
tor such as the mean and the weighted mean (c.f.
equations 2 and 3).
3.2.3 Content Based Filtering (CBF)
In CBF, as presented in (Mooney and Roy, 2000),
(Balabanovi
´
c and Shoham, 1997) and (Pazzani,
1999), the recommended items are those having simi-
lar features to the ones that the user have already liked
or purchased. Consequently, the implemented CBF
variants persue the following generic algorithm:
• Compute features similarities ISim(x
c
, x
i
) be-
tween the candidate item x
c
and all the other items
x
i
based on their sets of features
→
f
i
.
• Select S
x
c
, the neighborhood of the k most similar
items to the candidate item x
c
.
• Estimate ˆr
ac
based on the ratings assigned by u
a
to items in S
x
c
(c.f. equations 2 and 3).
3.2.4 Demographic Filtering (DF)
DF relies on predicting ratings based on both users’
demographic and items’ features similarities. Several
variants of DF are detailed in (Krulwich, 1997) and
(Pazzani, 1999). The general pattern of the evaluated
DF variants is as follows:
• Compute features similarities ISim(x
c
, x
i
) be-
tween x
c
and all the other items x
i
as in CBF.
• Select S
x
c
, the set of the k most similar items to
the candidate item x
c
.
• Compute demographic similarities USim(u
a
, u
i
)
between the active user u
a
and all the other users
u
i
based on their demographic attributes
→
d
i
.
• Select S
u
a
, the set of the L most demographically
similar users to u
a
.
• Estimate ¯r
ac
by aggregating the ratings assigned
by users in S
u
a
to the items in S
x
c
.
3.2.5 The Proposed Frequentist Model (FM)
Our recommendation model predicts consumers’ in-
terests and purchases using a statistical methodology
based on a set of variables that can be easily collected
in e-commerce platforms (Haddad et al., 2012). First,
users and items are clustered into groups and cate-
gories. Then, the probability p(e
k
|u, x, q) of observing
an evaluation (i.e. rating) e
k
∈
{
e
1
, e
2
. . . , e
N
e
}
being
assigned by a user u belonging to the group g
u
to an
item x of the category c
x
in the context q is calculated
for each possible value of the rating variable e. Fi-
nally, the ratings probabilities are used as an input to
estimate the rating ˆr
ux
that the user u would assign to
an unknown item x. In our model, we assume that
the users’ demographics and behaviors are induced
by their respective groups and that items’ attributes
depend only on their categories. Besides, we assume
that users’ interests modeled by the rating variable e
are induced by their groups, the products’ categories
and the consumption context. Those hypothesis de-
fine the conditional dependencies (and independen-
cies) between our model’s variables and enable us to
develop and simplify the rating probability as follows:
p(e
k
|u, x, q) = p(e
k
|g
u
, c
x
, q)p(g
u
|u)p(c
x
|x) (1)
p(c
i
|x) (resp. p(g
j
|u)) represent the membership
degree of an item x (resp. a user u) to the category
ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence
438
c
i
(resp. to the group g
j
). p(c
i
|x) and p(g
j
|u) are re-
lated to users and items clustering. However, our first
experiments using fuzzy clustering have shown that
such probabilistic clustering approaches decrease the
model’s prediction quality (Haddad et al., 2012). Be-
sides, the term p(e
k
|g
u
, c
x
, q) is closely related to col-
laborative, content based and demographic filtering
approaches and unifies their main ideas while adding
the contextual aspect.
3.3 Recommendation Engines
Several ratings estimators were implemented for the
recommendation algorithms in order to estimate the
rating r
ac
that the active user u
a
would assign to a
candidate item x
c
.
• Mean: let N be the number of existing ratings r
i, j
such as u
i
∈ S
u
a
andx
i
∈ S
x
c
. The mean estimator
is as follows:
ˆr
ac
=
1
N
∑
u
i
∈S
u
a
∑
x
i
∈S
x
c
r
i, j
(2)
• Weighted mean: ratings may be weighted using
USim(u
a
, u
i
), the similarity of a user u
i
to the ac-
tive user u
a
and/or ISim(x
c
, x
i
), the similarity of
an item x
i
to the candidate item x
c
. In DF, both
similarities are used (c.f. equation 3).
ˆr
ac
=
∑
u
i
∈S
u
a
∑
x
j
∈S
x
c
USim(u
a
, u
i
) · ISim(x
c
, x
j
) · r
i, j
∑
u
i
∈S
u
a
∑
x
j
∈S
x
c
USim(u
a
, u
i
) · ISim(x
c
, x
j
)
(3)
In UC-CF, items similarities are not taken into
account which leads to considering S
x
c
=
{
x
c
}
and
ISim(x
c
, x
c
) = 1 in equation 3. Analogically, in
IC-CF and CBF, we consider that S
u
a
=
{
u
a
}
and
USim(u
a
, u
a
) = 1.
In the proposed frequentist model (FM), once
users and items are clustered, we used the calculated
rating probabilities p(e
k
|u, x, q) as an input to the rat-
ing estimators. Hereby, the estimator calculates the
expected value of the rating variable such as
ˆr
ux
= E[p(e
k
|u, x, q)] =
∑
e
k
· p(e
k
|g, c, q) (4)
3.4 Clustering Methodologies
The clustering step of our recommendation approach
was carried using k-means and Expectation maxi-
mization techniques in order to build items categories
and users groups (Kaufman and Rousseeuw, 1990).
In order to cluster items and users based on the avail-
able data, several methodologies were used each feed-
ing different set of variables to the clustering algo-
rithm. Items were clustered using one of the follow-
ing methodologies:
1. Features clustering : clusteing items based on
their features leads to homogeneous clusters. This
favors the recommendation of items that are simi-
lar to the ones the user already appreciated.
2. Ratings clustering : using only items’ ratings for
clustering leads to clusters containing diverse but
correlated items as in IC-CF. This favors the rec-
ommendation and the discovery of new relevant
items with different features.
3. Mixed clustering : using features and ratings for
clustering leads to recommending items that are
similar or correlated to the ones that the user likes.
Similarly to items categories, users groups were
generated using three clustering methodologies:
1. Demographic clustering: using only users’ demo-
graphic attributes for clustering favors the recom-
mendation of items that are popular within a given
demographic class of users.
2. Ratings clustering: using only users’ ratings
for clustering helps regrouping like-minded users
with similar interests without necessarily being
similar demographically.
3. Mixed clustering: using both users’ demograph-
ics and assigned ratings for clustering helps re-
grouping users that are demographically similar
and/or having similar rating patterns. This favors
recommending not only demographically interest-
ing items, but also the ones that are correlated to
the user’s interests.
Several similarity and distance measures were im-
plemented for clustering such as Euclidian distance,
cosine, adjusted cosine, mean absolute error, root
mean square error and Manhattan distance.
4 EXPERIMENTATION
In order to evaluate the performances of our recom-
mendation model and assess the genericity of our rec-
ommendation architecture, we implemented several
state of the art recommendation approaches and con-
ducted a series of experimentations on a real world
dataset. In the following, the experimentations details
and results are presented and discussed.
4.1 Dataset
In this work, we adopted the MovieLens dataset since
it is the closest to our requirements. In fact, this
dataset includes ratings assigned by a set of users to
the movies they have watched. The data also includes
TowardsaGenericArchitectureforRecommendersBenchmarking
439
users’ demographics in addition to items’ features.
The dataset includes the following data:
• 1700 users : each user is described by his identi-
fier, age, gender and occupation (cf. table 1).
• 950 films : each movie is descibed by its identi-
fier, title, release date and the genres it belongs to
among the 19 predefined ones (e.g. Action, Ani-
mation, etc. . . ). Table 2 represents a data sample
describing a set of movies. The i
th
binary value of
the last column in table 2 is 1 if the corresponding
movie belongs to the i
th
predefined genre.
• 100000 ratings assigned by users to the available
movies with their respective timestamps. Hereby,
each user has at least 20 ratings.
Table 1: Users data sample.
ID Age Gender Occupation
1 24 M technician
3 23 M writer
Table 2: Items data sample.
ID Title Genres
22 Braveheart 0100000010000000010
23 Taxi Driver 0000000010000000100
29 Batman Forever 0110011000000000000
4.2 Experimental Results and
Benchmarking
Several configurations were evaluated for each ap-
proach using different combinations of similarity
measures and ratings estimators. Figure 2 aggregates
the obtained mean average error values for each ap-
proach in a box-and-whisker plot in order to display
the variability of its recommendation quality. The
bottom and top of each box indicate the upper and
the lower quartiles (respectively the 0.25 and 0.75
quantiles) of the obtained MAE values for a given ap-
proach, whereas, the band inside the box indicates the
median (the 0.5 quantile). Outlier values outside the
upper and lower whiskers are plotted with dots.
Table 3 enumerates the configurations of the pro-
posed recommendation model giving the best recom-
mendations with regard to the MAE and the RMSE
measures. Meanwhile, table 4 regroups the best re-
sults obtained using the candidate algorithms.
Our experiments show that UC-CF is the least sta-
ble approach since it depends heavily on its underly-
ing parameters, similarity measure and rating estima-
tor. However, the results obtained from demographic
filtering are similar and show that it is less sensi-
tive to its configuration. Meanwhile, item-centered
collaborative filtering, content-based filtering and our
frequentist model present similar prediction quality
ranges for the experimented configurations.
Item-centered collaborative filtering and content
based filtering present the results that are the clos-
est to ours. Although our recommendation model is
better with regard to MAE measure, IC-CF and CBF
have better results when considering RMSE as a qual-
ity measure. IC-CF’s results depend mainly on the
neighborhood size k determining the number of items
considered as similar to a candidate one. In fact, when
considering a small neighborhood of similar items,
IC-CF is able to recommend highly targeted but not
varied items which reflects a higher precision and a
lower recall. Besides, even if a larger neighborhood
of similar items leads to better balance between the
quality of the recommended items and their novelty,
such configuration would decrease the algorithms re-
sponsiveness due to the number of ratings to be ag-
gregated in order to make a prediction.
Content based filtering makes predictions based
on items similarities. The similarity measure and the
rating estimator have less influence on the algorithm’s
prediction quality. Similarly to IC-CF, our experi-
ments show that CBF responsiveness relies heavily
on the chosen neighborhood size. However, this pa-
rameter do not significantly influence the algorithm’s
prediction quality.
Demographic filtering shows similar results inde-
pendently of the employed similarity measure, rat-
ing estimator. However, predictions quality depends
mainly on items and users neighborhood sizes which
need to be fine tuned in a way that maximizes perfor-
mances and predictions quality.
The presented experimentations show that the pro-
posed recommender model captures consumers’ pur-
chase behaviors and is able to predict them. In fact, by
unifying the main ideas and hypothesis of the exper-
imented state of the art algorithms, our model is able
to predict and quantify consumers’ interests whether
if they are driven by demographics, items’ features or
by the context.
5 CONCLUSION AND FUTURE
WORKS
In this work, we proposed a generic architecture for
recommender systems development and benchmark-
ing including common recommendation algorithms,
clustering techniques and similarity measures. Using
the proposed architecture, we compared our recom-
mendation model with several existing algorithms and
showed that it encompasses their main ideas and may
ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence
440
Figure 2: Recommendation algorithms benchmarking results.
Table 3: Quality of the proposed recommendation model.
N
G
N
C
Users clustering Items clustering Similarity measure MAE RMSE
15 30 Demographics and ratings Ratings Euclidean distance 0.8145 1.0856
15 35 Demographics and ratings Ratings Euclidean distance 0.8452 1.0539
15 48 Demographics and ratings Features and ratings Euclidean distance 0.8195 1.0908
30 25 Ratings Ratings Euclidean distance 0.8171 1.0930
Table 4: Quality of state of the art algorithms.
Algorithm Neighborhood size Similarity measure Ratings estimator MAE RMSE
UC-CF 150 Adjusted Cosine Mean 0.9445 1.1842
UC-CF 150 Cosine Mean 0.9264 1.1712
UC-CF 150 Adjusted Cosine Weighted Mean 0.8784 1.1075
UC-CF 150 Cosine Weighted Mean 0.8674 1.0981
IC-CF 300 Euclidean Mean 0.8325 1.0432
IC-CF 300 Adjusted cosine Mean 0.8330 1.0434
CBF 300 Euclidean Mean 0.8325 1.0432
CBF 300 Euclidean Weighted Mean 0.8318 1.0446
DF users: 150 , items: 10 Cosine Weighted Mean 0.8862 1.1594
DF users: 150 , items: 10 Adjusted Cosine Weighted Mean 0.8939 1.1486
DF users: 300 , items: 5 Manhattan Weighted Mean 0.9235 1.1225
outperform their performances.
indent Similarly to the existing approaches, our model
is not fitted for consumption behaviors prediction in
static markets where purchases are driven by periodic
needs instead of compulsive interests. In fact, a sub-
set of consumers purchases are recurrent and often
include the same items (i.e. groceries, food, tv se-
ries, phone plans, etc. . . ). In this context, we pro-
posed a recommendation model for predicting recur-
rent and periodic consumption behaviors whose goal
TowardsaGenericArchitectureforRecommendersBenchmarking
441
is to complement the model presented in this paper
(Haddad et al., 2014). Future work will focus on ag-
gregating the two propositions into a more unified rec-
ommendation model and on releasing the framework
as an open source project in order to facilitate the
benchmarking of other recommendation approaches.
REFERENCES
Balabanovi
´
c, M. and Shoham, Y. (1997). Fab: content-
based, collaborative recommendation. Communica-
tions of the ACM, 40(3):66–72.
Berkovsky, S., Kuflik, T., and Ricci, F. (2008). Mediation
of user models for enhanced personalization in recom-
mender systems. User Model. User-Adapt. Interact,
18(3):245–286.
Bezdek, J. C. (1981). Pattern Recognition with Fuzzy
Objective Function Algorithms. Plenum Press, New
York.
Boutemedjet, S. and Ziou, D. (2008). A graphical model for
context-aware visual content recommendation. IEEE
Transactions on Multimedia, 10(1):52–62.
Burke, R. D. (2007). Hybrid web recommender systems. In
Brusilovsky, P., Kobsa, A., and Nejdl, W., editors, The
Adaptive Web, Methods and Strategies of Web Person-
alization, volume 4321 of Lecture Notes in Computer
Science, pages 377–408. Springer.
Deshpande, M. and Karypis, G. (2004). Item-based top-
N recommendation algorithms. ACM Transactions on
Information Systems, 22(1):143–177.
Goldberg, D., Nichols, D., Oki, B. M., and Terry, D. (1992).
Using collaborative filtering to weave an information
tapestry. Communications of the ACM, 35(12):61–70.
Haddad, M. R., Baazaoui, H., Ziou, D., and Ben Ghezala,
H. (2012). Towards a new model for context-aware
recommendation. In 6th IEEE International Confer-
ence Intelligent Systems (IS), Sofia, Bulgaria, pages
021 –027.
Haddad, M. R., Baazaoui, H., Ziou, D., and Ghezala, H. B.
(2014). A predictive model for recurrent consumption
behavior: An application on phone calls. Knowledge-
Based Systems, 64(0):32 – 43.
Herlocker, L., J., Konstan, A., J., 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,
Theoretical Models, pages 230–237.
Iyengar, S. S. and Lepper, M. R. (2000). When choice
is demotivating: can one desire too much of a good
thing? Journal of Personality and Social Psychology,
79(6):995–1006.
Jacoby, J., Speller, D. E., and Berning, C. A. K. (1974).
Brand choice behavior as a function of information
load: Replication and extension. Journal of Consumer
Research: An Interdisciplinary Quarterly, 1(1):33–
42.
Kaufman, L. and Rousseeuw, P. J. (1990). Finding Groups
in Data: an Introduction to Cluster Analysis. Wiley.
Krulwich, B. (1997). Lifestyle finder: Intelligent user pro-
filing using large-scale demographic data. AI Maga-
zine, 18(2):37–45.
Linden, G., Smith, B., and York, J. (2003). Amazon.com
recommendations: Item-to-item collaborative filter-
ing. IEEE Internet Computing, 7(1):76–80.
Mooney, R. J. and Roy, L. (2000). Content-based book rec-
ommending using learning for text categorization. In
Proceedings of the Fifth ACM Conference on Digital
Libraries, pages 195–204. ACM Press, New York.
Pazzani, M. J. (1999). A framework for collaborative,
content-based and demographic filtering. Artif. Intell.
Rev, 13(5-6):393–408.
Resnick, P., Iacovou, N., Suchak, M., Bergstorm, P., and
Riedl, J. (1994). Grouplens: An open architecture for
collaborative filtering of netnews. In Proceedings of
ACM 1994 Conference on Computer Supported Co-
operative Work, pages 175–186, Chapel Hill, North
Carolina. ACM.
Sarwar, B. M., Karypis, G., Konstan, J. A., and Riedl, J.
(2001). Item-based collaborative filtering recommen-
dation algorithms. In WWW, pages 285–295.
Schwartz, B. (2005). The paradox of choice: Why more is
less. Harper Perennial.
Woerndl, W. and Groh, G. (2007). Utilizing physical
and social context to improve recommender systems.
In Web Intelligence/IAT Workshops, pages 123–128.
IEEE.
ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence
442
