User Semantic Model for Dependent Attributes to Enhance Collaborative

Filtering

Sonia Ben Ticha

1,2

, Azim Roussanaly

, Anne Boyer

and Khaled Bsa

ıes

LORIA-KIWI Team, Lorraine University, Nancy, France

LIPA Lab, Tunis El Manar University, Tunis, Tunisia

Keywords:

Hybrid Recommender System, Latent Semantic Analysis, Rocchio Algorithm.

Abstract:

Recommender system provides relevant items to users from huge catalogue. Collaborative ﬁltering and

content-based ﬁltering are the most widely used techniques in personalized recommender systems. Collab-

orative ﬁltering uses only the user-ratings data to make predictions, while content-based ﬁltering relies on

semantic information of items for recommendation. Hybrid recommendation system combines the two tech-

niques. The aim of this work is to introduce a new approach for semantically enhanced collaborative ﬁltering.

Many works have addressed this problem by proposing hybrid solutions. In this paper, we present another

hybridization technique that predicts users preferences for items based on their inferred preferences for se-

mantic information of items. For this, we design a new user semantic model by using Rocchio algorithm and

we apply a latent semantic analysis to reduce the dimension of data. Applying our approach to real data, the

MoviesLens 1M dataset, signiﬁcant improvement can be noticed compared to usage only approach, and hybrid

algorithm.

1 INTRODUCTION

Recommender Systems (RS) provide relevant items

to users from a large number of choices. Several

recommendations techniques exist in the literature.

Among these techniques, there are those that provide

personalized recommendations by deﬁning a proﬁle

for each user. In this work, we are interested in

personalized recommender systems where the user

model is based on an analysis of usage. This model is

usually described by a user-item ratings matrix, which

is extremely sparse (≥ 90% of missing data).

Collaborative Filtering (CF) and Content-Based

(CB) ﬁltering are the most widely used techniques

in RS. The fundamental assumption of CF is that if

users X and Y rate n items similarly and hence will

rate or act on other items similarly (Su and Khoshgof-

taar, 2009). CB ﬁltering assumes that each user op-

erates independently and user will be recommended

items similar to the ones he preferred in the past (Lops

et al., 2011). The major difference between CF and

CB recommender systems is that CF uses only the

user-item ratings data to make predictions and recom-

mendations, while CB relies on item content (seman-

tic information) for recommendations. However, CF

and CB techniques must face many challenges like

the data sparsity problem, the scalability problem for

large data with the increasing numbers of users and

items.

To overcome the disadvantages of both techniques

and beneﬁt from their strengths, hybrid solutions have

emerged. In this paper, we present a new approach

taking into account the semantic information of items

in a CF process. In our approach, we design a new

hybridization technique, which predicts user prefer-

ences for items based on their inferred preferences for

latent item content; and presents a solution to the spar-

sity and scalability problems. Our system consists of

two components: the ﬁrst builds a new user model,

the user semantic model, by inferring user preferences

for item content; the second computes predictions and

provides recommendations by using the user seman-

tic model in a user-based CF algorithm (Resnick et al.,

1994) to calculate the similarity between users. The

originality of this work is in the building of the user

semantic model. Indeed, assuming that items are rep-

resented by structured data in which each item is de-

scribed by a same set of attributes, we build a user

semantic attribute model for each relevant attribute.

With this aim, we deﬁne two classes of attributes: de-

pendent and non dependent and we propose a suited

algorithm for each class. User semantic model is

205

Ben Ticha S., Roussanaly A., Boyer A. and Bsaïes K..

User Semantic Model for Dependent Attributes to Enhance Collaborative Filtering.

DOI: 10.5220/0004951102050212

In Proceedings of the 10th International Conference on Web Information Systems and Technologies (WEBIST-2014), pages 205-212

ISBN: 978-989-758-024-6

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

then deducted from the horizontal concatenation of

all user semantic attribute model. In previous works

(Ben Ticha et al., 2012; Ben Ticha et al., 2011) we

have presented solutions based on machine learning

algorithm to build a user semantic attribute model for

non dependent attribute. In this work, we present a

new approach for building a user semantic attribute

model for dependent attribute by using Rocchio al-

gorithm (Rocchio, 1971). Due to the high number

of attribute values, and to reduce the expensiveness

of user similarity computing, we apply a Latent Se-

mantic Analysis (LSA)(Dumais, 2004) to reduce the

size of the user semantic attribute model. We compare

our results to the standards user-based CF, item-based

CF and hybrid algorithms. Our approach results in an

overall improvement in prediction accuracy.

The rest of paper is organized as follows: Section

2 summarizes the related work. User semantic model

is described in Section 3. Section 4 describes our ap-

proach to build user semantic attribute model for non

dependent attribute. Section 5 describes the recom-

mendation component of our system. Experimental

results are presented and discussed in Section 6. Fi-

nally, we conclude with a summary of our ﬁndings

and some directions for future work.

2 RELATED WORK

Recommender System (RS) have become an indepen-

dent research area in the middle 1990s. CF is the

most widespread used technique in RS, it was the sub-

ject of several researches (Resnick et al., 1994; Sar-

war et al., 2001). In CF, user will be recommended

items that people with similar tastes and preferences

liked in the past (Adomavicius and Tuzhilin, 2005).

CB is another important technique; it uses techniques

developed in information ﬁltering research (Pazzani

and Billsus, 2007). CB assumes that each user oper-

ates independently and recommends items similar to

the ones he preferred in the past. Hybrid approach

consists on combining CF and CB techniques. The

Fab System (Balabanovic and Shoham, 1997) counts

among the ﬁrst hybrid RS. Many systems have been

developed since (Burke, 2007). Most of these hy-

brid systems do not distinguish between attributes

and treat their values in a same way. Moreover, be-

cause of the huge number of items and users, calcu-

lating the similarity between users in CF algorithm

became very expensive in time computing. Dimen-

sion reduction of data is one of the solution to re-

duce the expensiveness of users similarity comput-

ing. Mobasher et al. (Mobasher et al., 2003) com-

bine values of all attributes and then apply a LSA to

reduce dimension of data. Sen et al. (Sen et al., 2009)

are inferring user preferences for only one attribute,

the item’ tags, without reducing dimension. Manzato

(Manzato, 2012) computes a user semantic model for

only the movie genre attribute and applies a Singular

Value Decomposition (SVD) to reduce the dimension

of data. In our approach, we compute a user semantic

attribute model for each relevant attribute and we ap-

ply a suited reduction dimension algorithm for each

attribute class.

3 USER SEMANTIC MODEL

In this paper, we are interested only to items described

by structured data. According to the deﬁnition of Paz-

zani et al. (Pazzani and Billsus, 2007), in structured

representation, item can be represented by a small

number of attributes, and there is a known set of val-

ues that each attribute may have, for instance, the at-

tributes of a movie can be title, genre, actor and di-

rector. In the following, we will use the term feature

to refer to an attribute value, for instance Documen-

tary, Musical and Thriller are features of movie genre

attribute.

3.1 Dependent and Non Dependent

Attribute

In structured representation, each attribute has a set of

restricted features. However, the number of features

can be related or not to the number of items. That is

why we have deﬁned two classes of attributes:

• Dependent Attribute: attribute, which having

very variable number of features. This number is

closely related to the number of items. So, when

the number of items is increasing, the number of

features is increasing also. For example: directors

and actors of movies, user tags.

• Non Dependent Attribute: attribute, which hav-

ing a very few variable number of features, and

this number is not related to the number of items.

Thus, the increasing number of items has no effect

on the number of features. For example: movie

genre, movie origin and cuisine of restaurants.

In addition, all attributes do not have the same

degrees of importance to users. There are attributes

more relevant than others. For instance, the movie

genre can be more signiﬁcant, in the evaluation crite-

ria of user, than the origin. Experiments that we have

conducted (see Section 6.2) conﬁrmed this hypothe-

sis. In this paper, we assume that relevant attributes

will be provided by a human expert. Therefore, for

WEBIST2014-InternationalConferenceonWebInformationSystemsandTechnologies

206

each relevant attribute A, we build a user semantic at-

tribute model that predicts the users preferences for

its features (or group of features). This model is de-

scribed by a matrix Q

(users in lines and features (or

group of features) of A in columns). In our approach,

we design a suited algorithm for building the user se-

mantic attribute model for each class of attribute. For

non dependent attribute, due to the low number of fea-

tures, we have used a clustering algorithm. Section

3.2 brieﬂy described the operating principle of our

solution that have been addressed in previous works

(Ben Ticha et al., 2012; Ben Ticha et al., 2011). For

dependent attribute, we have explored techniques is-

sues from information retrieval (IR) research. Section

4 presents our solution for building the user semantic

attribute model for dependent attribute that is the aim

of this paper. The user semantic model for all relevant

attributes, described by the matrix Q, is the result of

the horizontal concatenation of all user semantic at-

tribute models Q

3.2 User Semantic Model for Non

Dependent Attribute

Let us denote by S the set of items, U the set of users, s

a given item ∈ S, u a given user ∈U and a rating value

r ∈

{

1,2,...,5

}

≡ R. U

the set of users that rating the

item s, then we deﬁne the rating function for item s

by δ

: u ∈ U

7−→ δ

(u) ∈ R. We denote also by F

the set of features of attribute A, f a given feature

∈ F

and S

the set of items associated to feature f .

For instance if we consider the movie genre attribute,

action

is the set of all action movies.

An item s is represented by its usage proﬁle vec-

tor s

= (δ

(u)−δ

)

(u=1..

)

, where δ

is the average

rating of all rated items by user u. The idea is to par-

tition all items described by their usage proﬁle in K

clusters, each cluster is labeled by a feature f ∈ F

(or a set of features).

The number K of clusters and the initial center of

each cluster is computed by the initialization step of

the clustering algorithm. In initial step, each clus-

ter C

consists of items in

f labeling C

and labeled

by the set of corresponding features; so its center is

the mean of its items described by their usage proﬁle

vector s

. Moreover, an attribute can be mono val-

ued or multivalued depending on the number of fea-

tures that can be associated to a given item s. For

example, the attribute movie genre is multivalued be-

cause a movie can have several genres while movie

origin is a mono valued attribute because a movie has

only one origin. Thus, if an attribute is multivalued,

s can belong to several clusters C

, while for mono

valued attribute, an item should belong only to one

cluster. Therefore, for multivalued attribute, the clus-

tering algorithm should provide non disjointed clus-

ters (a fuzzy clustering), whereas, for mono valued

attribute, the clustering algorithm should provide dis-

jointed clusters.

After running the clustering algorithm, we obtain

K cluster centers; each center k is described by a vec-

tor c

= (q

k,u

)

(u=1..

)

. The K centers is modeling

k latent variables issued from the features of the at-

tribute A. Thus, the user semantic attribute model is

described by the matrix Q

= (q

u,k

)

(u=1..

, k=1..K)

With non dependent attribute, the number of as-

sociated features is low, this is why the clustering is

suitable. Moreover, the user semantic attribute model

allows an important reduction of dimension and so re-

duce the expensiveness of user similarity computing.

In (Ben Ticha et al., 2011), we have used the Fuzzy

CMean Algorithm on the movie genre attribute, we

have obtained good performance because the user se-

mantic attribute model has no missing values and all

similarities between users were able to be computed.

In (Ben Ticha et al., 2012), we have used the KMean

clustering algorithm on the movie origin attribute.

Because of the missing values in the user item rating

matrix, we have proposed an algorithm for the initial-

ization step of the KMean clustering using a movie

origin ontology. We obtained good results compared

to user-based CF but not as good as results for the

genre attribute.

4 USER SEMANTIC MODEL FOR

DEPENDENTS ATTRIBUTES

For a dependent attribute A, the set F

of its features

can be important and it augments with the increasing

of the set of items S. In this paper, we present our

solution to compute a user semantic attribute model

for dependent attribute.

In addition to the formalism used in Section 3.2,

we denote by F

the set of features f ∈ F

associ-

ated to item s and by S

the set of items s ∈ S rated

by user u. We deﬁne also, the rating function of

user u as δ

: s ∈ S

7→ δ

(s) ∈ R; and the Item Fre-

quency Function for item s ∈ S as f req

: f ∈ F

7→

1 i f f ∈ F

( f associated to item s), 0 otherwise. The

Frequency Item Matrix F = ( f req

( f ))

s∈S and f ∈F

provided by computing f req

( f ) for all items and all

features.

UserSemanticModelforDependentAttributestoEnhanceCollaborativeFiltering

207

4.1 Computing the TF-IDF on the

Frequency Item Matrix F

One of the best-known measures for specifying key-

word weights in Information Retrieval is the TF-

IDF (Term Frequency/Inverse Document Frequency)

(Salton, 1989). It is a numerical statistic, which re-

ﬂects how important a word is to a document in a

corpus. In our case, we replace document by item

and term by feature and compute TF-IDF on the Fre-

quency Item Matrix F.

FF( f ,s) =

f req

( f )

max

f req

( j)

(1)

where the maximum is computed over the f req

( j) of

all features in F

of item s.

The measure of Inverse Document Frequency

(IDF) is usually deﬁned as:

IDF( f ) = log



(2)

where



is the number of items assigned to feature

f (ie f req

( f ) 6= 0). Thus, the FF-IDF weight of fea-

ture f for item s is deﬁned as:

ω(s, f ) = FF( f , s) × IUF( f ) (3)

4.2 Rocchio Formula for User Semantic

Attribute Model

Rocchio algorithm (Rocchio, 1971) is a relevance

feedback procedure, which is used in information re-

trieval. It designed to produce improved query for-

mulations following an initial retrieval operation. In

a vector processing environment both the stored in-

formation document D and the requests for informa-

tion R can be represented as t-dimensional vectors of

the form D = (d

,...,d

) and B = (b

,...,b

In each case, d

and b

represent the weight of term i

in D and B, respectively. A typical query-document

similarity measure can then be computed as the inner

product between corresponding vectors.

Rocchio showed in (Rocchio, 1971), that in a re-

trieval environment that uses inner product computa-

tions to assess the similarity between query and docu-

ment vectors, the best query leading to the retrieval of

many relevant items from a collection of documents

is:

opt

∑

−

∑

(4)

Where Di represent document vectors, and

is the

corresponding Euclidean vector length; R is the set of

relevant documents and NR is the set of non relevant

documents.

We use the Rocchio formula (4) for computing the

user semantic proﬁle of user u. In our case we replace

D by S the collection of items and term by feature.

Thus, the user semantic model Q

(u) for user u and

attribute A is equal to Q

opt

in formula (4). The set of

relevant items R for user u is composed of all items

in S having δ

(s) ≥ δ

. The set of non relevant items

NR for user u is composed of all items in S having

(s) < δ

4.3 Latent Semantic Analysis for

Dimension Reduction

Latent Semantic Analysis (LSA) (Dumais, 2004) is a

dimensionality reduction technique which is widely

used in information retrieval. Given a term-document

frequency matrix, LSA is used to decompose it into

two matrices of reduced dimensions and a diagonal

matrix of singular values. Each dimension in the re-

duced space is a latent factor representing groups of

highly correlated index terms. Here, we apply this

technique to create a reduced dimension space for the

user semantic attribute model. In fact, for depen-

dent attribute, the number of feature is correlated to

the number of items, and so it can be very elevated

and even higher than the number of items. Thus,

the semantic user attribute model can have dimension

greater than the user rating matrix thereby aggravat-

ing the scalability problem.

Singular Value Decomposition (SVD) is a well

known technique used in LSA to perform matrix de-

composition. In our case, we perform SVD on the

frequency item matrix F

by decomposing it into

three matrices:

F = I

∗ Σ

r,r

∗V

(5)

where I and V are two orthogonal matrices; r is the

rank of matrix F, and Σ is a diagonal matrix, where

its diagonal entries contain all singular values of ma-

trix F and are stored in decreasing order. I and V

matrices are the left and right singular vectors, corre-

sponding to item and feature vectors in our case. LSA

uses a truncated SVD, keeping only the k largest sin-

gular values and their associated vectors, so

= I

∗ Σ

∗V

(6)

is the rank-k approximation to F, and is what LSA

uses for its semantic space. The rows in I

are the item

vectors in LSA space and the rows in V , are the fea-

ture vectors in LSA space. In the resulting Frequency

Item Matrix, F

, each item is, thus, represented by a

set of k latent variables, instead of the original fea-

tures. This results in a much less sparse matrix, im-

proving the results of users similarity computations

WEBIST2014-InternationalConferenceonWebInformationSystemsandTechnologies

208

in CF process. Furthermore, the generated latent vari-

ables represent groups of highly correlated features in

the original data, thus potentially reducing the amount

of noise associated with the semantic information.

In summary, for building the user semantic at-

tribute matrix Q

for a dependent attribute A; ﬁrst,

we apply a TF-IDF measure on Frequency Item Ma-

trix F; second, we reduce the dimension of Frequency

Item Matrix F by applying a LSA; third, we compute

the user semantic attribute model by using Rocchio

formula (4).

5 RECOMMENDATION

To compute predictions for the active user u

, we use

the user-based CF algorithm (Resnick et al., 1994).

User-Based CF predicts the rating value of active user

on non rated item s ∈ S, it is based on the k-

Nearest-Neighbors algorithm. A subset of nearest

neighbors of u

are chosen based on their similarity

with him or her, and a weighted aggregate of their rat-

ings is used to generate predictions for u

. Equation

7 provides formula for computing predictions.

p(u

,s) = δ

+ L

∑

v∈V

sim(u

,v)(δ

(s) − δ

) (7)

where L =

∑

v∈V

sim(u

,v)

and V is the set of the nearest

neighbors (most similar users) to u

that have rated

item s. V can range anywhere from 1 to the number

of all users.

sim

(

u,v) =

∑

u,k

− q

)(q

v,k

− q

)

∑

u,k

− q

)

∑

v,k

− q

)

(8)

The function sim(u,v) provides the similarity be-

tween users u and v and is computed by using

the Pearson Correlation (8). In the standard user-

based CF algorithm, the users-items rating matrix

(δ

(s)

(u∈U, s∈S)

) is used to compute users’ similarities.

In our algorithm, for computing the similarities be-

tween users we use instead the user semantic matrix

Q. As we have already mentioned, the matrix Q is

the horizontal concatenation of user semantic attribute

model Q

for each relevant attribute A.

Although we apply a user-based CF for rec-

ommendation, our approach is also a model-based

method because it is based on a new user model to

provide ratings of active user on non rated items. Our

approach resolves the scalability problem for several

reasons. First, the building process of user seman-

tic model is fully parallelizable (because the comput-

ing of user semantic attribute model is done in in-

dependent way for each other) and can be done off

line. Second, this model allows a dimension reduc-

tion since the number of columns in the user seman-

tic model is much lower than those of user item rat-

ing matrix, so, the computing of similarities between

users is less expensive than in the standard user-based

CF. In addition, our approach allows inferring simi-

larity between two users even when they have any co-

rated items because the users-semantic matrix has less

missing values than user item ratings matrix. Thus,

our approach provides solution to the neighbor transi-

tivity problem emanates from the sparse nature of the

underlying data sets. In this problem, users with sim-

ilar preferences may not be identiﬁed as such if they

haven’t any items rated in common.

6 PERFORMANCE STUDY

In this section, we study the performance of our ap-

proach, User Semantic Collaborative Filtering (USCF

in plots), against the standards CF algorithms: User-

Based CF(UBCF) (Resnick et al., 1994), and Item-

Based CF(IBCF) (Sarwar et al., 2001) and an hybrid

algorithm. We evaluate these algorithms in terms of

predictions accuracy by using the Mean Absolute Er-

ror (MAE) (Herlocker et al., 2004), which is the most

widely used metric in CF research literature. It com-

putes the average of the absolute difference between

the predictions and true ratings in the test data set,

lower the MAE is, better is the prediction.

We have experimented our approach on real data

from the MovieLens1M dataset of the MovieLens

recommender system

. The MovieLens1M provides

the usage data set and contains 1,000,209 explicit rat-

ings of approximately 3,900 movies made by 6,040

users. For the semantic information of items, we

use the HetRec 2011 dataset (HetRec2011, 2011) that

links the movies of MovieLens dataset with their cor-

responding web pages at Internet Movie Database

(IMDb) and Rotten Tomatoes movie review systems.

We use movie genre and movie origin as non depen-

dent attributes, movie director and movie actor as de-

pendent attributes.

We have ﬁltered the data by maintaining only

users with at least 20 ratings, and available features

for all movies. After the ﬁltering process, we obtain

a data set with 6020 users, 3552 movies, 19 genres,

44 origins, 1825 directors and 4237 actors. The usage

data set has been sorted by the timestamps, in ascend-

ing order, and has been divided into a training set (in-

cluding the ﬁrst 80% of all ratings) and a test set (the

last 20% of all ratings). Thus, ratings of each user in

http://www.movielens.org

UserSemanticModelforDependentAttributestoEnhanceCollaborativeFiltering

209

(a) (b) (c)

Figure 1: Impact of LSA on prediction accuracy of Rocchio algorithm.

test set have been assigned after those of training set.

It should be noted that the building of user semantic

attribute model for the non dependent attributes genre

and origin have been addressed respectively in previ-

ous works (Ben Ticha et al., 2011; Ben Ticha et al.,

2012). Therefore, we will not detail the experiments

conducted for these attributes in this paper. If it is not

speciﬁed, the number of nearest neighbors is equal to

60.

6.1 Impact of LSA on Prediction

Accuracy

In Figure 1, the MAE has been plotted with respect

to the LSA rank. It compares the Rocchio approach

with and without applying LSA (dimension reduc-

tion) on director attribute (Figure 1(a)), actor attribute

(Figure 1(b)) and combined attribute director actor

(Figure 1(c)). In all cases, the plots have the same

look, the MAE of Rocchio with LSA decreases until

it reaches the MAE value of Rocchio without LSA.

So, LSA dimension reduction has no effect on im-

proving the accuracy. This can be explained by the

fact that features are not highly correlated, which is

understandable especially for attributes director and

actor, hence their poor performance. Indeed, for the

director attribute, for instance, the MAE without re-

duction (1825 features) is equal to 0.7122 while the

best value with LSA is equal to 0.7884. However, for

combined attributes director actor (6062 features),

the best value is equal to 0.7083 (obtained for Rocchio

without LSA) while the worst value is equal to 0.7145

(Rocchio with LSA, rank=400). For rank equal to

1200, MAE= 0.7096, so a dimension reduction about

80% for a loss of accuracy about 0.18%. In this

case, features of combined attribute, actor

director,

are more correlated than the features of each attribute

taken alone hence, the best performance. Although

the LSA doesn

t improve the accuracy, dimension re-

duction is signiﬁcant. Thus, it allows to reduce the

cost of users similarity computing, specially when the

number of features is very high, as is the case of com-

bined attributes director actor.

6.2 Impact of Attribute Class on

Prediction Accuracy

Figure 2 compares algorithms for building user

semantic attribute model in term of MAE. The

Average algorithm (Average in plot) is build-

ing user semantic attribute model by computing

the average of user ratings by feature (q

(u, f )

AV G

{

(s)/s ∈ S

and f ∈ F

}

). Fuzzy C Mean al-

gorithm (FuzzyCM in plot) is a fuzzy clustering used

for non dependent and multivalued attribute (here

genre) and KMean algorithm (KMean in plot) is

used on non dependent and mono valued attribute

(here origin). Moreover, Rocchio algorithm (Roc-

chio in plot) is applied here for all attributes depen-

dent and non dependent. For genre, origin and di-

rector attributes, Rocchio without LSA provides best

results than with dimension reduction. For actor at-

tribute, LSA with rank equal to 1100 is applied (Roc-

chio+LSA in plot). When analyzing this ﬁgure we

note ﬁrst, that Average algorithm provides, for all at-

tributes, the worst performance compared to all other

algorithms. Second, if we applied the Rocchio al-

gorithm to non dependent attribute the performance

compares unfavorably against the dependent attribute,

while the best performance is attained by FuzzyCM

algorithm on genre attribute and the difference is im-

portant (0.7079 for FuzzyCM and 0.7274 for Roc-

chio). This allows to deduct that, using a suited al-

gorithm for each attribute class provides best perfor-

mance than applying the same algorithm for all at-

tributes. Third, the origin attribute has the worst per-

formance compared to the other three attributes and

this for all algorithms; this is conﬁrm our hypothe-

WEBIST2014-InternationalConferenceonWebInformationSystemsandTechnologies

210

Figure 2: Impact of user semantic attribute algorithm on prediction accuracy.

Figure 3: Evaluation of USCF against standards CF.

sis that all attributes don’t have the same relevance to

users. The attribute origin can be less signiﬁcant in

the choice of users than the genre, actor or director,

which is intuitively understandable.

6.3 Comparative Results of USCF

against Standard CF Systems

Figure 3 depicts the recommendation accuracy of

User Semantic Collaborative Filtering (USCF) in

contrast to standard Item-Based CF (IBCF) and

User-Based CF (UBCF). USCF-<Attributes> in plot

means the list of relevant attributes involved in build-

ing the user semantic model Q. For each relevant

attribute, the suited algorithm is applied. So, Fuzzy

CMean for genre, KMean for origin, and Rocchio

with LSA (rank=1200) for combined attribute direc-

tor actor. Furthermore, MAE has been plotted with

respect to the number of neighbors (similar users) in

the k-nearest-neighbor algorithm. In all cases, the

MAE converges between 60 and 70 neighbors. Our

approach, USCF (in plot) results in an overall im-

provement in accuracy for all attributes. In addition,

the best performance is achieved by the combination

genre-origin-director actor. This improvement can

be explained by many reasons. First, taking into ac-

count the semantic proﬁle of items in a CF recommen-

dation process. Second, for non dependent attribute,

user semantic model is built according to a collabora-

tive principle; ratings of all users are used to compute

the semantic proﬁle of each user. It is not the case of

the Average algorithm; this may explain its results de-

spite taking into account the semantic aspect. Third,

the choice of the attribute can have signiﬁcant inﬂu-

ence on improving the accuracy. Lastly, users seman-

UserSemanticModelforDependentAttributestoEnhanceCollaborativeFiltering

211

tic model Q has few missing values, so, it allows in-

ferring similarity between two given users even when

they have any items rated in common.

7 CONCLUSION AND FUTURE

WORK

The approach presented in this paper is a compo-

nent of a global work, which the aim, is to seman-

tically enhanced collaborative Filtering recommenda-

tion and to resolve the scalability problem by reducing

the dimension. For this purpose, we have designed

a new hybridization technique, which predicts users’

preferences for items based on their inferred prefer-

ences for semantic information. We have deﬁned two

classes of attributes: dependent and non dependent at-

tribute, and presented a suited algorithm for each class

for building user semantic attribute model. The aim

of this paper is to present our approach for building

user semantic attribute model for dependent attribute.

We have deﬁned an algorithm based on Rocchio al-

gorithm and have applied Latent Semantic Analysis

(LSA) for dimension reduction. Our approach pro-

vides solutions to the scalability problem, and alle-

viates the data sparsity problem by reducing the di-

mensionality of data. The experimental results show

that USCF algorithm improves the prediction accu-

racy compared to usage only approach (UBCF and

IBCF) and hybrid algorithm (Average). In addition,

we have shown that applying Rocchio formula on non

dependent attribute, decreases signiﬁcantly the pre-

diction accuracy compared to results obtained with

machine learning algorithms. Furthermore, we have

experimentally shown that all attributes don’t have the

same importance to users. Finally, experiments have

shown that the combination of relevant attributes en-

hances the recommendations.

An interesting area of future work is to use ma-

chine learning techniques to infer relevant attributes.

We will also study the extension of the user seman-

tic model to non structured data in witch items are

described by free text. Lastly, study how our ap-

proach can provide solution to the cold start prob-

lem in which new user has few ratings. Indeed, CF

cannot provide recommendation because similarities

with others users cannot be computed.

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. IEEE Trans.

Knowl. Data Eng., 17(6):734–749.

Balabanovic, M. and Shoham, Y. (1997). Fab: Content-

based, collaborative recommendation. Commun.

ACM, 40(3):66–72.

Ben Ticha, S., Roussanaly, A., and Boyer, A. (2011).

User semantic model for hybrid recommender sys-

tems. In The 1st Int. Conf. on Social Eco-Informatics

- SOTICS, Barcelona, Espagne. IARIA.

Ben Ticha, S., Roussanaly, A., Boyer, A., and Bsa

ıes, K.

(2012). User semantic preferences for collaborative

recommendations. In 13th Int. Conf. on E-Commerce

and Web Technologies - EC-Web, pages 203–211, Vi-

enna, Austria. Springer.

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

The Adaptive Web, volume 4321 of Lecture Notes in

Computer Science, pages 377–408. Springer.

Dumais, S. T. (2004). Latent semantic analysis. An-

nual Review of Information Science and Technology,

38(1):188–230.

Herlocker, J. L., Konstan, J. A., Terveen, L. G., and Riedl,

J. (2004). Evaluating collaborative ﬁltering recom-

mender systems. ACM Trans. Inf. Syst., 22(1):5–53.

HetRec2011 (2011). In 2nd Int Workshop on Information

Heterogeneity and Fusion in Recommender Systems.

The 5th ACM Conf. RecSys.

Lops, P., de Gemmis, M., and Semeraro, G. (2011).

Content-based recommender systems: State of the

art and trends. In Recommender Systems Handbook,

pages 73–105. Springer US.

Manzato, M. G. (2012). Discovering latent factors from

movies genres for enhanced recommendation. In The

6th ACM conf. on Recommender systems - RecSys,

pages 249–252, Dublin, Ireland.

Mobasher, B., Jin, X., and Zhou, Y. (2003). Semantically

enhanced collaborative ﬁltering on the web. In 1st Eu-

ropean Web Mining Forum, volume 3209, pages 57–

76, Cavtat-Dubrovnik, Croatia.

Pazzani, M. and Billsus, D. (2007). Content-based recom-

mendation systems. In The Adaptive Web, volume

4321 of Lecture Notes in Computer Science, pages

325–341. Springer Berlin Heidelberg.

Resnick, P., Iacovou, N., Suchak, M., Bergstrom, P., and

Riedl, J. (1994). Grouplens: an open architecture for

collaborative ﬁltering of netnews. In The 1994 ACM

conf. on Computer supported cooperative work, pages

175–186, Chapel Hill, North Carolina, USA.

Rocchio, J. (1971). Relevance feedback in information re-

trieval. In Salton, G., editor, The Smart Retrieval Sys-

tem - Experiments in Automatic Document Process-

ing, chapter 14, pages 313–323. Prentice-Hall, Inc.

Salton, G. (1989). Automatic Text Processing. Addison-

Wesley.

Sarwar, B. M., Karypis, G., Konstan, J. A., and Riedl, J.

(2001). Item-based collaborative ﬁltering recommen-

dation algorithms. In The 10 Int. WWW Conf., pages

285–295, Hong Kong, China.

Sen, S., Vig, J., and Riedl, J. (2009). Tagommenders: con-

necting users to items through tags. In The 18th Int

Conf. on WWW, pages 671–680, Madrid, Spain.

Su, X. and Khoshgoftaar, T. M. (2009). A survey of collabo-

rative ﬁltering techniques. Adv. Artiﬁcial Intellegence.

WEBIST2014-InternationalConferenceonWebInformationSystemsandTechnologies

212