A Recommendation System for Enhancing the Personalized Search
Itineraries in the Public Transportation Domain
Aroua Essayeh and Mourad Abed
Univ. Lille Nord de France, F-59000 Lille, France
UVHC, LAMIH, F-59313 Valenciennes, France
CNRS, UMR 8201, F-59313 Valenciennes, France
Keywords: Clustering, Learning Machine, Ontology, Public Transportation.
Abstract: In traditional transport information systems, the users must explicitly provide the information related to both
their profiles and travels to receive a personalized response. However, this requires, among others, an extra
effort from user in term of search time. We aim to identify not only implicitly users’ information, but also to
anticipate their need even if some data are missing through a recommender system based on collaborative
filtering technique. In this work, the information related to users is represented using the ontology which
proved far more adequate model for representing semantically data.
1 INTRODUCTION
In the era of globalization, the development of the
intelligent systems has seen a rapid evolution due to
the emergence of new technologies. However, the
information systems suffer nowadays from the
proliferation of information. The information is very
heterogeneous and provided from various sources.
For this purpose, personalization plays a decisive role
in information systems. In public transportation field,
the main difficulty is to propose for a user the best
itinerary that fit with his preferences and profile with
reducing at the same time both search time and effort.
Moreover, most of these systems do not tackle with
the problem of missing information, in this case, the
user is forced to fill and to complete all needed
information. Furthermore, it will be more
sophisticated to prognosticate their requirements
instead of expressing them at every turn. Hence, this
will save time when browsing for the best result, and
will help novice user when finding their needs more
easily.
To this end, we propose a new strategy of
personalization based on reasoning on over two
ontologies related to user profile and transport
domain , we have to learn later from their past
interaction with the system in order to reformulate the
query and recommend a new personalized solution.
The paper has three main contributions. First, we
propose to model users’ profile by using the ontology
and we include some properties that we considered
important in the context of travel. Second, we use
jointly, inferences rules and fuzzy clustering
algorithm to anticipate user’s needs implicitly even if
some information seems important are missed
according to the stored histories. Third, this algorithm
is enhanced through a new dissimilarity measure to
handle the problem of heterogeneous data.
The rest of this paper is organized as follows: We
introduce first a background about the personalized
information in public transport field. Second, we
discuss the collaborative filtering approaches. Third,
we announce our motivation in the next section.
Then, we illustrate the different steps of our proposed
algorithm. Finally, we show the experimental results
and evaluations, we discuss some conclusions about
the benefits and limitations of our approach and we
outline some future works.
2 RELATED WORKS
2.1 Personalized Information System in
Public Transport
As defined in (Hagen 1999), personalization is “The
ability to provide content and services that are
tailored to individuals based on knowledge about
their preferences and behaviour”. In public
transportation field, the personalization has a
Essayeh, A. and Abed, M.
A Recommendation System for Enhancing the Personalized Search Itineraries in the Public Transportation Domain.
DOI: 10.5220/0006315904150423
In Proceedings of the 19th International Conference on Enterprise Information Systems (ICEIS 2017) - Volume 1, pages 415-423
ISBN: 978-989-758-247-9
Copyright © 2017 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
415
significant consequence when searching for
personalized itineraries. This itinerary must, among
others, respond to users’ preferences and needs. Many
approaches have dealt with the personalization in this
field. An MDA (Model Driven Architecture)
approach is applied in (Marçal de Oliveira et al. 2013)
in order to build a personalized itinerary for a user
giving him at the same time some services related to
his preferences. The weakness of this approach is in
the manual mapping used between the domain
ontology and the context model which let this system
not applicable in real cases and remain a theoretic
approach. In their study, (Moussa, Soui2 & Abed
2013) have introduced a multi-criteria decision
making approach to personalize a system in public
transportation. This method uses the ELECTRE
method; it focuses on achieving a compromise
between the different compensatory criteria.
However, this work considers some quantitative
information related to the travel and assumes
implicitly that all criteria are considered fully
comparable which is not always possible in complex
systems. In their recent work, (Bouhana et al. 2015)
have proposed a hybrid method based on CBR (Case
Based Reasoning) and ontology to personalize the
itinerary for stakeholders. Despite the efficiency of
the adopted methodology, some drawbacks are still
unresolved, among which the prediction of user’s
needs without involving their personal motivations.
2.2 Collaborative Filtering Approach
Besides the huge number of heterogeneous data, users
have some difficulties to express clearly their needs
in a significant timing. In practice, users need
recommendations because they do not have enough
knowledge to make an autonomous decision (Ricci ,
Rokach & Shapira 2015) or what is the response
relative to their request.
Collaborative filtering techniques aim to perform
personalized recommender system. Furthermore,
several works are found in the literature and have
investigated to identify recommender system that aim
to provide a personalized content for a target user. As
defined in (Liu et al. 2014), the recommender
system’s aims to “guess” the users’ preferences by
analysing their behaviours when interacting with the
system. In other words, it reveals historic users to
anticipate their needs. Nonetheless, such systems
suffer from both scalability, and the cold start
problem. This issue is addressed in the present paper.
Moreover, these systems require the determination of
the correlation between the user and the
recommended service. This latter is then another
challenge to mitigate. We identify the most known
techniques used to deal with this issue such as the
Pearson correlation technique (Resnick, Iacovou &
Suchak 1994), the constrained Pearson correlation
(CPC)-based similarity (Shardanand & Maes 1997),
the cosine-based similarity (Sarwar et al. 2000), and
the adjusted cosine based measures (Ahn 2008) to
identify the similar users that rated the same services.
Some approaches require aggregator model to
enhance decision system. For example, a combination
of the OWA (Ordered Weight Aggregator) and the
LSP aggregator (Logic Scoring of Preference) used in
(Moreno et al. 2013) to analyse the user’s interactions
with the system, and to identify the degree of interests
to promote decision making. Choquet integral is also
used in both (Bouhana et al. 2015) and (Bouhana et
al. 2013) for making a decision. AHP and OWA
aggregators are used in (Abolghasem Sadeghi &
Kyehyun 2009) to investigate the decision maker.
The major key of a personalized process is to
know what a user wants and also why he needs this
(Lakiotaki & Matsatsinis 2011). To deal with this
issue a single rating item for each item doesn’t offer
an appropriate understanding as it use a unique
criteria to predict his needs. This issue is addressed
by the use of a multi-criteria approach (Nilashi , bin
Ibrahim & Ithnin 2014) (Liu, Mehandjiev & Xu 2011)
(Lakiotaki & Matsatsinis 2011). These ones can
provide more information about the user
requirements.
To that extent, we need to build communities that
gather users who shared common interests. We
reason over these communities to derive similar
behaviours and provide recommendations. In this
light, clustering techniques have not only the goal to
gather users into several groups, but also to construct
communities to learn and to recommend latter similar
solution. Without any prior knowledge, we aim to
obtain overlapping clusters, and a single user would
belong to more than one cluster. With hard
classification, user is associated only to one cluster,
while the soft classification allows a user to join many
clusters with different degrees of membership. The
most known method is the fuzzy K-means algorithm.
Sometimes, it is combined with others data mining
algorithms such as ANN (Artificial Neural Networks)
presented in (Paireekreng & Wai Wong n.d.) in order
to classify users respect to their demographics data
and interests. The key drawback of this research study
is that it treated with very limited resources relating
to the mobile application (Paireekreng & Wai Wong
n.d.). (Lazzerini & Marcelloni 2007) have applied a
method gained on an unsupervised algorithm namely
the Fuzzy Divisive Hierarchical Clustering (UFDHC)
ICEIS 2017 - 19th International Conference on Enterprise Information Systems
416
algorithm to classify the users of a web portal into
groups with similar characteristics and interests
represented by prototype. This research emphasizes
only the content of web pages related to specific
profiles. The Competitive Agglomeration for
Relational Data (CARD) algorithm is used in (Gandy
et al. 2005) to classify user’s session. (Castellano et
al. 2007)have focused on web personalization
problem and especially on discovering what current
webpage related to user profile. The clusters are
generated by mining the log data of a web containing
user’s preferences. Moreover, (Teran & Meier 2010)
have proposed a fuzzy recommender system for
election field. Their work had to identify the most
similar candidates according to voter’s preferences
and tendencies. They presented a modified fuzzy c-
means algorithm. A similar work, (Jalali et al. 2010)
have defined an architecture based on two phases to
predict the user future requests. The first phase is
turned on offline mode and it implemented the
clustering model for navigation pattern mining; it
consisted in computing the degree of connectivity
between each pair of the Web pages, and then created
an undirected graph to find the connect component.
The online phase has to predict the user future
intentions through mining Web server logs.
Table 1: Comparison between the fuzzy clustering
approaches.
Approaches Techniques Data type
(Paireekreng &
Wai Wong n.d.)
K-means and
ANN algorithm
Categorical
data
(Lazzerini &
Marcelloni 2007)
Fuzzy Divisive
Hierarchical
Clustering
(UFDHC)
characteristics
and interests
(Gandy et al.
2005)
CARD algorithm
Relational
Data
(Martin-Bautista
et al. 2002)
fuzzy
classification
rules
Categorical
data
(Teran & Meier
2010)
fuzzy c-means
algorithm
Categorical
data
(Jalali et al. 2010)
clustering model
with indirect
graph
Categorical
and numerical
data
The major challenge to consider with the use of fuzzy
clustering algorithm is to distinguish, as much as
possible, between the inter-clustering in order to
obtain a best differentiated subset, and to minimize as
much as possible the intra-class inertia, with the aim
of obtaining the most homogeneous possible clusters.
2.3 Motivation
There is no doubt that personalized systems have
been gaining interest in many domains and especially
in the transport field. Despite the various methods and
techniques presented in the literature, these
approaches have weaknesses and limitations. The
response must not only reply to user’s request, but
also it must anticipate his expectations before even he
expresses them. In this setting, we intend to explore
the histories of users to, one hand learns from their
experiences and on the other hand, to handle with the
problem of explicit or missed information. To this
end, the techniques of machine leaning seem be the
most appropriate to resolve this issue. We aim then to
gather users who have common interests to learn from
their experience. For this purpose, we opt for using
the technique of fuzzy clustering and we propose a
new dissimilarity measure to tackle the problem of
heterogeneous data.
3 GENERAL PROPOSED
FRAMEWORK
In this section, we describe the general proposed
framework for enhancing personalized research. The
aim of this proposal is to recommend for a current
user the most similar response to his request. In this
light, we split our proposal into two parts; the first
concerns the modelling phase and the second part
aims to respond to the user’s request according to our
recommender based model.
The figure 1 describes the input and the output of
the proposed reasoning process.
Figure 1: Reasoning process.
A Recommendation System for Enhancing the Personalized Search Itineraries in the Public Transportation Domain
417
3.1 User Profile Modelling
This section presents the characteristics of the user
profile ontology. The ontology was implemented
using Protégé beta 5.0 [25] in OWL DL and its
consistency tested using Jena and OWL Reasoner.
The user profile ontology was created in order to
facilitate the extraction of the user personal
information, needs, and interests, under the context of
personalization.
In this light, the user profile is defined as the
union of:

∪
∪
∪

With:
P
s
: represents his personal data (age, gender,
address, ability, proficiency, user’s state).
P
p
: depicts the user’s preferences for a precise
service. The preferences may be related to the travel
(cost, duration, walking, correspondence) or to the
user’s personal requirements (accommodation,
administration, entertainment and healthcare).
P
EX
: defines the users’ histories (past query and
validated choice).
P
c
: defines contextual elements related to the user
such as time and location.
Figure 2: Top-level hierarchy of Ontology User Profile.
The figure 3 describes the main concepts of this
ontology, the properties (represent the attributes) and
the objects properties (used to establish relationships
between classes)
Figure 3: Classes, dataproperties, objectproperties,
description.
3.2 Reasoning based Process
In this section, we have many challenges to
overcome; First, unlike others approaches which used
a few criteria to respond to user’s request, we aim in
this paper to compensate different types of criteria to
obtain a personalized solution (itinerary). Further,
these criteria are generally treated as numerical data,
whereas it can be expressed in different forms. Even
more, users’ profiles are also different, and results
change from a profile to another. In this view, we
build our learning process. Accordingly, we propose
the fuzzy k-mode algorithm (Huang & K. Ng 1999)
within the inference engine to deal with the issues
mentioned above.
3.2.1 Initialization of Clustering Parameters
We symbolize then the terminologies used in our
clustering algorithm:
X,Y are two different users, 


,

,

,….,

represent the classes related to
user profile, xi={ x
1
,x
2
,..,x
N
}: instances related to the
first user and y
i
={y
1
,y
2
,…,y
N
} are the instances
related to the second user, Cluster: is a group of users
who share common characteristics (history, profile,
context and preference), Reference Vector (V
ref
)represents the various concepts used in the input of
our algorithm, Vector mode is specific to each cluster
ICEIS 2017 - 19th International Conference on Enterprise Information Systems
418
and composed of the instances related to the concepts
of reference vector, Dissimilarity measure represents
the degree of closeness between two individuals,
Concept criterion( C
crit
)represents the concept topic
in each cluster, Concept non-criterion (C
Ncrit
)
represents the other concepts which may influence the
clustering process, Fuzzy index α, and finally, K is
the number of the cluster. The fuzzy clustering
algorithm is executed on 4 four step: 1) Selection
modes, 2) Computing of dissimilarity measure, 3)
Computing of membership matrix and 4) Upgrading
clusters. In the next section, we outline the different
steps of this algorithm.
3.2.2 Selection Vector Mode
The most straightforward approach is to choose
arbitrarily k random modes, but it turns out that it is
the weakest point of the conventional clustering
algorithm; the fuzzy k-mode algorithm as an example
(Choia & Chungb 2017). In addition, the correct
choice of k is often ambiguous. For this purpose,
several research efforts investigated the issue to more
wisely choose this parameter as discussed in
(Kodinariya & Makwana 2013). For the initialization
process, we choose k according to the theorem of rule
of thumb which is also similar to heuristic rule
(Madhulatha 2012). Thereupon, k is calculated as
following (n is the number of users) Eq.1

/2
(1)
First, the reference vector is composed of classes
from user profile. The choice of vector mode depends
on this vector, and it defines the instances related to
each class. For this purpose, we calculate the
frequency of instances using the following Eq2.
F
r
(xi)=

(2)
With n
xi
: is the number of users that have the same
instance in the cluster and N is the total number of
users.
The selection of vector modes has two main
conditions; first, the instances of each vector mode
must not be duplicated in the other one and second, in
case of missing values, each empty instance must
follow our proposition of normalization method
explained in detail later with Eq.4
3.2.3 New Dissimilarity Measure
In order to deal with different type of data, we
propose a new dissimilarity measure. Further, the
comparison between each two instances (xi; yi) is
1
http://www.tfidf.com/
calculated based on their similarities and the type of
their related concepts (Ci (xi); Ci (yi)). We consider
for two individuals X and Y, the same reference
vector. Therefore, as we have different types of data
(numeric, non-numeric (textual and semantic)),
We assume that the concept criteria and non-
criteria are determined according to their weight in
the database as discussed in the next paragraph.
Forthwith, we propose four discussed cases:
Case1: if (Ci (xi); Ci (yi)) C
crit
and xi = yi xi
and yi are two similar instances of the same
concept (non-numeric). Their related concept is
considered as a concept criterion. Example: Ci (xi)
= Ci (yi) = "FamilyStatus"; xi = yi = "single".
Case2: if (Ci (xi), Ci (yi)) C
Ncrit
and xi # yi xi
and yi are two similar instances of the same
concept but this concept does not belong to the
concept criterion. Example: Ci (xi) = Ci (yi) =
"gender"; xi = "female" and yi = "male".
Case3: if (Ci (xi); Ci (yi)) {C
crit
; C
Ncrit
} and (xi;
yi) are numeric values xi and yi belong to the
same concept. Example: Ci (xi) = Ci (yi) = "age";
xi = "13" and yi = "20".
Case4: if (Ci (xi); Ci (yi)) C
crit
and xi and yi
are two different instances but belong to the same
concept. This latter belongs to the concept
criterion. Example: Ci (xi) = Ci (yi) =
"PreferenceTravel"; xi = "fast" and yi = "cheaper".
The enhanced similarity measure is then:
,




,


∶1

,


:2

,
∈
,


,

:3




,


∶4
0




(3)
The similarity is calculated between two users’
vector. Therefore, for case1 and case 2, we opt for
using Jaccard similarity (Jeff M. 2013) for computing
similarities between two strings or textual attributes,
each attributes is associated to a weight. For this, we
use the technique of TF-IDF
1
. With this method, we
can define which concept (class) is a concept
criterion. Each instance is associated with a weight w
i
to evaluate how important this instances in database.
For case3, we compute the similarity using the
method proposed in (Bouhana et al. 2013) for
numerical attributes. Afterwards, OWA aggregator
(Yager 1988) is applied to compute the average
similarity. The weights are generated automatically
according to the orness measure and dispersion
measure. The case 4 discusses uncertain criteria. For
this propose, a
WordNet
2
is applied to search the
synonyms, the hypernym, the hyponym and any
2
https://wordnet.princeton.edu/
A Recommendation System for Enhancing the Personalized Search Itineraries in the Public Transportation Domain
419
existing relation between the attributes of each users.
However, in order to deal with missing information,
a normalization technique is adopted. We assign a
value (ε) for each empty attribute, we have then:

0



1 


(4)
As a result, the distance D is defined then as the ratio
of the sum of similarities measures between the
attributes related to each individual and the sum of θ
i
.
The new formulation is given below:
,

,


(5)
3.2.4 Membership Matrix Upgrading
The next step of our fuzzy clustering algorithm is to
calculate and update the membership matrix
[

]. The membership matrix allows the degree
of user closeness with its corresponding cluster to be
identified. This value is to be updated as far as we do
not hit the stop condition.
The formula is expressed as follows:


1 


0


1
1





 





(6)
Where
X
i
: is the set of user’s attributes, C
i
: represents the
current mode concepts of a cluster which a user
belongs to and C
k
: represents the concepts of others
modes which a user does not belong to.
3.2.5 User’s Clustering Upgrading
The mode related to each cluster is not static since the
algorithm did not reach the stop condition. To update
the mode, we calculate the following formula:
ω

∗

ω

(7)
Where:
ω

is the membership matrix for the ith iteration α is
the fuzziness index and X
i
is the instances of user X.
This step continuous and the new mode is compared
with the previous one until attending the stop
condition. The stop condition is obtained if and only
if the following objective function is minimized:
m
,

∑∑



,
(8)
3.2.6 Recommendation Process
Hereinafter, we describe the execution of our
recommendation process; the algorithm is described
as follows:
Input
Mod
0
: initial vector mode
X={x
1
, x
2
,…., x
n
}: instances of user X
= fuzzifier index
N=number of clusters
t =1 // number of iterations
ω

: Membership matrix
Z

: Vector mode upgraded
Output
Clusters= {CL
,CL
,CL
,…,CL
} //set of
clusters
Begin
//Initialize the fuzzifier
index
N=Numbers_of_concepts_criterion
//Construction of modes
Foreach CL
in clusters do
Foreach C
crit
(x)
Begin
1) Initialize the vector mode
(see Eq.(1 and 2))
2) Calculate new similarity
measure (see Eq.(3))
3) Calculate new dissimilarity
measure
,
(see
Eq.(5))
If
,
is the less one then
do
XCluster
i
// save user
profile in cluster
End If
4) Calculate ω

based on Z

with
Eq.(6)
t++
5) Calculate

based
on

withEq.(7)
6) //Verify objective function
(seeEq.(8))
If (

,

=F
(

,

thenStop.
Else
Go step (7)
End if
7) Determine

based on

If F (

,

=F
(

,

thenStop.
Else Go step (4)
End if
8) Repeat Step (4) to step
(7), until there is no
movement between clusters
(minimize the objective
function)
End
End
Return similar user in
current cluster
Return recommended solution
End
ICEIS 2017 - 19th International Conference on Enterprise Information Systems
420
4 EXPERIMENTAL
EVALUATION
In this section, we focus on the following three
questions. One is to check whether such initialled
parameters (fuzziness index and number of clusters)
negatively affect the quality of resulting clusters
(final solution). The other is to experimentally
evaluate how our clustering algorithm is in agreement
with other alternative algorithms and finally, we
evaluate the proposed solution by addressing the
following research question. How accurately the
proposed approach can provide pertinent solution?
This is addressed by calculating the precision and the
recall. For this setting, a case study is presented in the
next section.
4.1 Case Study
Personalized information systems in intelligent
public transportation domain are complex systems
and dealing with a large amount of heterogeneous
data from various sources. These systems must
respond to user’s queries based on their requirements
and also profiles. In fact, in traditional systems, user
should provide all needed information explicitly to
get a personalized response. However, user interacts
with the system in order to search a personalized
itinerary even if he doesn’t dispose of all the travel’s
information; modality, time-tables and tariffs. In this
setting, we recommend for him an itinerary in a
multimodal network based on the deducted
community of similar users. For the instance, we are
interested only on the recommendation process. We
consider a community of 14 users, every user
executes 3 different queries, and we have at least 42
solutions. Based on the inferred knowledge provided
by this community, we will recommend for the
current user the most suitable itinerary.
To this end, ontology of the domain that describes
the field of transport has been exploited and enriched
by RATP OPEN DATA
3
and GTFS
4
. The figure 4
describes an extract of this ontology
3
https://opendata.paris.fr/explore/
Figure 4: Classes, dataproperties, objectproperties,
description.
4.2 Evaluation Results and Discussion
We have implemented our method in Java language
and executed in Intel core i3 processor (1.9GHz) with
2GB memory running on windows 7 operating
system.
Clustering Quality
In order to justify the quality of clustering results, we
use the Cohen’s kappa value, which is widely used
especially to measure the agreement between the
proposed algorithm and the other competitors. The
values between -1 and 1 assert that the accord is low,
or very low, moderate agreement or strong
agreement. The formula applied with Kappa value is:


1
(9)
With P
0
: is the relative observed agreement among
ratters and P
e
is the hypothetical probability of chance
agreement, using the observed data to calculate the
probabilities of each observer randomly saying each
category.
Table 2: Cohen’s Kappa.
Results
New fuzzy
K-modes
Clustering
Fuzzy
K-means
clustering
K-modes
clustering
New fuzzy
K-modes Clustering
1
0,67
0,71
Fuzzy
K-means clustering
0,67
1 0,4
K-modes clustering 0,71 0,4 1
4
https://developers.google.com/transit/gtfs/reference/
A Recommendation System for Enhancing the Personalized Search Itineraries in the Public Transportation Domain
421
As shown in this table, the Kappa value is calculated
for each pair of clustering techniques used. The new
fuzzy K-modes clustering and the fuzzy K-means
have good agreement strength of 0,67 (0,67
0,6;0,80, while new Fuzzy K-modes and K-modes
clustering have an agreement strength of 0,71.
We admit that our proposed algorithm seems to
show good results.
Precision, Recall and F-measure
The Precision is defined by dividing the number of
users correctly belonging to the positive cluster by the
total number of users belonging to the positive cluster
while Recall is calculated as the ratio between the
number of users correctly to the positive cluster and
the total number of elements that actually belong to
the positive cluster. Indeed, the results is a set of
clusters, we evaluate these measures by calculating
the average of the precision and the recall in each
clusters. Finally, we compute the F-measure for each
query.



,


,2
∗

Table 3: Evaluation results.
Query
Estimated
solution
Recom
mended
Solution
Precision Recall
F-
measure
Q1
S8, S7,
S27, S38
1) S12, S27,
S42 2) S8,
S15, S23,
S30, S38
3) S14
0,33 0,75 0,46
Q2
S28, S13,
S33
1) S13, S33,
S28 2) S11,
S26, S35,
S37 3) S6,
S21
0,33 1 0,50
Q3
S8, S28, S1,
S30
1) S8, S15 2)
S5 3) S13,
S18 4) S20,
S23, S30,
S38
0,22 0,5 0,31
Figure 5: Precision, Recall and F-measure evolutions.
The found results presented in Fig. 5 shows that our
algorithm returned most of the relevant results
according to the high recall's values.
5 CONCLUSIONS
In this paper, we propose a new recommendation
system based on fuzzy clustering algorithm jointly the
ontologies in the public transportation field. The
proposal supports both qualitative and quantitative
data and aims to gather users who share common
features into the same cluster. By building such
clusters, called communities, we raise the problem of
explicit information by learning from the similar
profiles according to their interactions’ histories with
the system. The recommender solution fit his needs
and responds to his formulated request. In our future
work, we will detail how we reformulate the query
sending by the user by adding the new implicit
information inferred from the proposed clustering and
we manage to use BabelNet to overcome the limits of
the WordNet.
REFERENCES
Abolghasem Sadeghi, N & Kyehyun, K 2009, 'Ontology
based personalized route planning system using a multi-
criteria decision making approach', vol 36, no. 2250–2259.
Ahn, HJ 2008, 'A new similarity measure for collaborative
filtering to alleviate the new user cold-starting
problem', vol Volume 178, Issue 1, 2 January 2008,
Pages 37–51.
Bouhana , , Zidi, A, Chabchoub, & Abed, M 2015, 'An
ontology-based CBR approach for personalized
itinerary search systems for sustainable urban freight
transport'.
ICEIS 2017 - 19th International Conference on Enterprise Information Systems
422
Bouhana, , Fekih, A, Abed, M & Chabchoub, H 2013, 'An
integrated case-based reasoning approach for
personalized itinerary search in multimodal
transportation systems'.
Castellano, , Mesto, , Minunno, M & Alessandra Torsello,
2007, 'Web User Profiling Using Fuzzy Clustering',
2007.
Choia, D-W & Chungb, C-W 2017, 'A K-partitioning
algorithm for clustering large-scale spatio-textual data',
Information Systems, pp. Information Systems 64
(2017) 1–11.
Gandy, L, Dept. of Comput. Sci., SIU,CIU, Rahimi, S &
Gupta, B 2005, 'A modified competitive agglomeration
for relational data algorithm', Fuzzy Information
Processing Society, 2005. NAFIPS 2005. Annual
Meeting of the North American, IEEE.
Hagen, 1999, 'mart Personalization,” Forrester Report'.
Huang , Z & K. Ng, M 1999, 'A Fuzzy k-Modes Algorithm
for Clustering Categorical Data', Journal IEEE
Transactions on Fuzzy Systems, vol Volume 7 Issue 4,
August 1999 Page 446-452.
Jalali, M, Mustapha, N, Sulaiman, MN & Mamat, 2010,
'WebPUM: A Web-based recommendation system to
predict user future movements', vol Volume 37, Issue
9, September 2010, Pages 6201–6212.
Jeff M. , P 2013, 'Jaccard Similarity and Shingling',
University of Utah.
Kodinariya, TM & Makwana, PR 2013, 'Review on
determining number of Cluster in K-Means Clustering',
vol Volume 1, Issue 6, November 2013, no. ISSN:
2321-7782.
Lakiotaki, & Matsatsinis, NF 2011, 'Multi-Criteria User
Modeling in Recommender Systems', vol Vol26, no. 64
- 76.
Lazzerini, B & Marcelloni, F 2007, 'A hierarchical fuzzy
clustering-based system to create user profiles'.
Liu , H, Hu, , Mian, A, Tian, H & Zhu, X 2014, 'A new user
similarity model to improve the accuracy of
collaborative filtering', vol 56 (2014) 156–166.
Liu, , Mehandjiev, & Xu, D-L 2011, 'Multi-Criteria Service
Recommendation Based on User Criteria Preferences'.
Madhulatha, TS 2012, 'An overview on clustering
methods,', CoRR abs/1205.1117.
Marçal de Oliveira, K, Bacha, F, Mnasser, H & Abed, M
2013, 'Transportation ontology definition and
application for the content personalization of user
interfaces', Expert Systems with Applications ,
www.elsevier.com/locate/eswa, pp. 3145-3159.
Martin-Bautista, MJ, Kraft, DH, Vila, MA, Chen, J & Cruz,
J 2002, 'User profiles and fuzzy logic for web retrieval
issues'.
Moreno , A, Valls , A, Isern , D, Marin, L & Borras, J 2013,
'SigTur/E-Destination:Ontology-based personalized
recommendation of Tourism and Leisure Activities',
vol 26(2013)633–651.
Moussa, , Soui2, M & Abed, 2013, 'User Profile and Multi-
Criteria Decision Making: Personalization of
Traveller’s Information in Public Transportation', 17th
International Conference in Knowledge Based and
Intelligent Information and Engineering Systems -,
Procedia Computer Science 22 ( 2013 ) 411 – 420.
Nilashi , M, bin Ibrahim, & Ithnin, N 2014, 'Multi-criteria
collaborative filtering with high accuracy using higher
order singular value decomposition and Neuro-Fuzzy
system'.
Paireekreng, W & Wai Wong, K, 'Intelligent Mobile User
Profile Classification forContent Personalisation'.
Resnick, , Iacovou, N & Suchak, M 1994, 'GroupLens: An
Open Architecture for Collaborative Filtering of
Netnews', no. pp. 175–186.
Ricci , , Rokach, L & Shapira, 2015, 'Introduction to
Recommender Systems Handbook'.
Sarwar, , Karypis, G, Konstan, & Riedl, 2000, 'Item-based
Collaborative Filtering Recommendation Algorithms',
viewed 26 May 2016,
<http://www10.org/cdrom/papers/519/node12.html>.
Shardanand, U & Maes, 1997, 'Social information filtering:
algorithms for automating word of mouth,'.
Teran, & Meier, 2010, 'A Fuzzy Recommender System for
eElections'.
Teran, & Meier, 2010, 'A Fuzzy Recommender System for
eElections', Electronic Government and the
Information Systems Perspective, First International
Conference, EGOVIS 2010, Bilbao, Spain, August 31 -
September 2, 2010. Proceedings.
Yager, RR 1988, 'On ordered weighted averaging
aggregation operators in multi-criteria decision
making', IEEE Transactions on Systems, Man and
Cybernetics, vol volume 18 Issue 1, January/February
1988 , pp. Pages 183 - 190.
A Recommendation System for Enhancing the Personalized Search Itineraries in the Public Transportation Domain
423