Ahmad Hawalah and Maria Fasli
Department of Computer Science and Electronic Engineering, University of Essex, Colchester, U.K.
Keywords: User profile, Ontology, Mapping, Concept clustering, Web personalisation.
Abstract: Web personalization systems that have emerged in recent years enhance the retrieval process based on each
user’s interests and preferences. A key feature in developing an effective web personalization system is to
build and model user profiles accurately. In this paper, we propose an approach that implicitly tracks users’
browsing behaviour in order to build an ontology-based user profile. The main goal of this paper is to
investigate techniques to improve the accuracy of this user profile. We focus in particular on the mapping
process which involves mapping the collected web pages the user has visited to concepts in a reference
ontology. For this purpose, we introduce two techniques to enhance the mapping process: one that maintains
the user’s general and specific interests without the user’s involvement, and one that exploits browsing and
search contexts. We evaluate the factors that impact the overall performance of both techniques and show
that our techniques improve the overall accuracy of the user profile.
Although the Internet and the WWW have enhanced
access to information, their rapid expansion have
also caused information overload to such an extent
that the process of finding a specific piece of
information or a suitable product may often become
frustrating and time-consuming for users. One way
to deal with this problem is through adaptive or
personalization web systems (Pignotti, Edwards and
Grimnes, 2004, Challam, Gauch and Chandramouli,
2007, Sieg, Mobasher and Burke, 2007 and Pan,
Wang and Gu, 2007). The ultimate objective of
these systems is to provide personalized services or
products with respect to each user’s requirements.
Today, the use of personalization systems is
widespread in many application domains. For
example, in the domain of e-learning,
personalization has been used to provide each user
with specific information that meets his or her needs
and knowledge (Azouaou and Desmoulins, 2007). In
the e-commerce domain, a personalization system
plays an important role in recommending products
or services based on the user’s needs and interests;
for instance, when a user navigates through a
specific section in a retail book store, the system can
recommend books that suit his or her characteristics
and preferences (Gorgoglione, Palmisano &
Tuzhilin, 2006).
Of course, all of these systems require some
information about users in order to learn and
respond to their interests and needs. Each system
independently models and builds a user’s profile,
which is a representation of known information
about that individual, including demographic data,
interests, preferences, goals and previous history.
However, one of the main challenges in current
personalization systems is that they rely mostly on
static or low-level dynamic user profiles (Felden and
Linden, 2007 and Trajkova and Gaunch, 2004)
which constrain the personalization process because
they use the same user information over time, often
leading to recommendations of irrelevant services as
the user’s needs and interests change.
One way to overcome this challenge is by
building an ontological user profile that dynamically
captures and learns the user’s interests and
preferences (Challam, Gauch and Chandramouli,
2007, Daoud, Tamine and Boughanem, 2008,
Middleton et al. 2004 and Felden and Linden, 2007).
The process of building such a profile is complex
and requires multiple tasks. These tasks can be
divided into three main phases: the information
retrieval (IR) phase, which consists of preparing the
reference ontology, collecting user navigation
Hawalah A. and Fasli M..
DOI: 10.5220/0003179003210328
In Proceedings of the 3rd International Conference on Agents and Artificial Intelligence (ICAART-2011), pages 321-328
ISBN: 978-989-8425-40-9
2011 SCITEPRESS (Science and Technology Publications, Lda.)
behaviour, and mapping URLs to the reference
ontology; the profile adaptation and learning phase,
which forms the dynamic user profile; and the
personalization engine phase, which provides
recommendations based on the dynamic user profile.
In this paper, we investigate the first phase, mapping
URLs to a reference ontology, which is essential for
the subsequent phases. Indeed, capturing inaccurate
user interests in the first phase would directly affect
the personalization performance. Therefore, this
paper gives much needed attention to the first phase
by introducing two novel algorithms that aim to
improve the mapping process. These two algorithms
each have characteristics that enhance particular
aspects of the mapping process. The first algorithm,
called Gradual Extra weight (GEW), is applied to an
ontology to maintain a balance between a user’s
general and specific interests. The second algorithm,
called the Contextual Concept Clustering (3C), is
designed to exploit the user’s context and thereby
improve the mapping accuracy.
The paper is structured as follows. Next we
discuss related work and following that we discuss
the process of modelling the user profile. Section
four presents the details of the two techniques and
the next section presents a set of experiments that
have been conducted along with the results. The
paper ends with the conclusions and pointers to
future work.
A number of approaches have been presented to
improve the overall accuracy of the mapping
process. One such approach is to use a reference
ontology with a limited number of levels. Liu et al.
(2002), for example, mapped users’ interests to a
level-two ontology, while other approaches utilized
a three-level ontology to map users’ interests (Chen,
Chen and Sun, 2002 and Trajkova & Gauch, 2004).
As to retrieval precision, using a limited number of
levels has been reported to improve overall
accuracy, but a great limitation of this approach is
that levels two or three of the ontology may still be
too general to represent a user’s actual interests.
Another approach that has been applied to map
interests to an ontology is adding a pre-defined
percentage of each sub-concept’s weight to its super-
concept. The idea behind this approach is that if a
user is interested in a particular concept, then he or
she also has some interest in its more general super-
concept. Middleton et al. (2004) and Kim et al.
(2007) applied this approach by adding an extra 50%
for each concept’s weight to its super-concept’s
weight, and then repeating the process until the root.
Although this method offered an improvement over
the original cosine similarity approach, its
accumulation behaviour led to more emphasis on the
top-level concepts, which are too general to
represent a user’s actual interests, while the middle
and low-level concepts receive less attention.
Daoud, Tamine and Boughanem (2008), on the other
hand, assumed that representing interests with two
levels of the ontology is too general, while leaf-node
representation is too detailed, and that the most
relevant concept is the one that has the greatest
number of dependencies. Based on these
assumptions, they proposed a sub-concept
aggregation scheme, the main goal of which was to
represent all user interests with three levels in the
ontology. The weight of a level-three concept in this
system is calculated by adding the weights of all its
sub-concepts and then associating each user’s
interests to one level-three concept.
All of the approaches that have been proposed to
improve the mapping process have limitations. Users
tend to have general and specific interests on a wide
range of topics. Therefore, assuming a two or three-
level representation of all users' interests cannot be
accurate or particularly effective. For instance, in the
Open Directory Project (ODP) ontology, level-two
Computers/Programming and level-three
Computers/Programming/Languages are both too
general to represent, for example, an interest in Java
or C# programming languages. On the other hand,
approaches that rely on adding extra weight to a
super-concept based on its sub-concepts also suffer
from a serious limitation since the accumulation
behaviour leads to more emphasis on top-level
concepts, which are too general to represent actual
user interests. Therefore, we need a new approach
that is capable of maintaining both general and
specific interests. The focus of this paper is on
introducing new techniques that can maintain a
balance between general and specific interests.
In this section, we present our approach for
modelling ontological user profiles. The process of
modelling user profiles involves three aspects: (i)
tracking the user behaviour; (ii) using a reference
ontology; (iii) mapping concepts to the ontology.
These are explained in the subsequent sections.
ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence
3.1 Tracking User Behaviour
In order to learn and discover user interests, some
information about users is required. Since collecting
user data explicitly adds more burden on users (Kim
and Chan, 2003), in this system we aim at collecting
user browsing behaviour implicitly. The main data
that we need to observe in this system is the visited
websites and timestamp which denotes the date/time
at which a website is visited. For each session, the
user navigation behaviour is recorded and stored in a
log file. After each session, the contents of each
visited website are extracted. It is essential at this
point to remove all the noise by applying text
analysis techniques. Various algorithms are applied
like tokenization, sentence splitting and stemming
(Porter, 1980). After performing the text analysis,
the processed data are stored in the processed log
file (P-log file).
3.2 Using a Reference Ontology
Ontology representation is a rich knowledge
representation which has been proven to provide a
significant improvement in the performance of the
personalization systems (Trajkova and Gauch, 2004,
Challam, Gauch and Chandramouli, 2007, Daoud,
Tamine and Boughanem, 2008, Middleton et al.,
2004). In this paper, an ontology plays a key role in
modelling the user profile. A reference (or domain)
ontology provides a clear illustration of contents of a
particular domain of application (Trajkova and
Gauch, 2004). The ontology is modelled in a
hierarchal way in which super-concepts are linked to
sub-concepts. Another potential feature of using a
reference ontology is that it could be agreed and
shared between different systems, and therefore, user
interests and preferences which mapped to the
ontology can be easily shared between different
systems. Unlike flat representations, a reference
ontology provides a richer representation of
information in that semantic and structural
relationships are defined explicitly. In this paper,
user interests are generated from the reference
ontology based on the user browsing behaviour.
After each browsing session, the visited websites are
mapped to the reference ontology in order to classify
each webpage to the right concept. A vector space
mechanism is used in this paper as the main
classifier (see equation 1).
Term weight= (
Despite the fact that the term frequency or vector
space classifier is one of the simplest classification
methods, it has a few drawbacks. One important
drawback is that this classifier distinguishes between
terms or vectors that have the same root. Words such
as “play”, “plays” and “played” are processed as
different words. This makes the classifier less
effective in that the dimensionality of the terms
increases. In order to reduce the dimensionality of
the terms, we use the Porter stemming algorithm to
remove term suffix and return each term to its stem
, 1980). Stop words also can be removed
from the reference ontology. Words such as “and”,
“he” and “but” add more noise to the classifier and
consequently lead the classifier to be less effective.
3.3 Mapping Concept to an Ontology
Once the term weights are calculated for each term
in the ontology, any vector similarity method can be
used to map visited web pages to appropriate
concepts (or classes) in the reference ontology. In
this paper, a cosine similarity algorithm (Baeza &
Ribeiro, 1999) which is a well known algorithm is
applied to classify websites to the right concepts.
In this section we introduce two novel approaches
that are capable of maintaining the user’s general
and specific interests without the user’s
4.1 Gradual Extra Weight (GEW)
The idea behind GEW is that if a user is interested in
a particular concept, then he also has some interest
in its super-concept. Unlike other approaches that
were discussed in section 2, in this approach we
make no assumption about the number of levels in
an ontology as the specification of each ontology
varies. Additionally, we do not assign a specific
percentage of a sub-concept to be added to its super-
concept. Instead, we propose an auto-tuning
mechanism in which the percentage value of each
sub-class that is added to its super-class is tailored to
different levels on the ontology (see equation 2).
Extra percentage (EP) = (CL/2)*10 (2)
CL: the current sub-class's level
In this approach, we assume that the concepts deep
in any ontology are more closely related than those
in higher levels. Therefore, the Extra Percentage in
our approach is calculated by dividing the level of a
sub-concept by two and then the result is multiplied
by 10 to give the extra percentage to be added from
the sub-concept to its super-concept. As we move up
towards the root, the percentage is reduced. In this
case, we keep a balance between the general and the
specific interests. Algorithm 1 describes the
procedure used to map and calculate the EP.
Input: reference ontology and web pages that need to be mapped
Output: URLs with top α concepts from the ontology with
updated similarity weights
RO= reference ontology
RO= {
}, concepts with associated documents.
= level of a concept c
LOG= log file that contains user's browsing behaviour
LOG ={
} visited web sites.
EP= Extra percentage= (CL /2)*10)
SR= Sim Results between URL and concepts after applying GEW.
// apply original Cosine similarity for each URL
Foreach 
LOG do
Extract contents;
Apply dimensionality reduction techniques;
Calculate TFIDF;
Foreach c
∈RO do
Calculate sim(
); // using cosine algorithm
);// add URL, concepts and sim weight.
// Select top α concepts to apply GEW on.
SR.sort by weight;
Foreach c
∈RS and c
.count do
If sim.weight > 0 then
Calculate EP= (c
Extra_weight= EP×c
.superclass-weight += Extra_weight;
SR.sort by weight; // re-order SR after applying GEW.
Algorithm 1: Gradual Extra Weight.
4.2 Contextual Concept Clustering
Though the GEW approach may improve the
process of mapping web pages to concepts, correct
mapping cannot be guaranteed as not all the visited
web pages usually have good representative
contents. Therefore, we further improve the mapping
process by taking advantage of having a log file that
stores the entire user’s browsing behaviour. Usually,
when users browse the Internet, they tend to visit
several web pages that represent one interest. We
call our mechanism Contextual Concept Clustering
(3C) because the context of the user behaviour is
considered. To illustrate, a visited web page could
be clustered to one concept in one session, but in
another session, the same web page could be grou-
ped to a different cluster. This behaviour could be
significant in the process of finding the right user
interests. Therefore, we introduce the 3C mechanism
that aims at grouping related web pages with the
same concept into one cluster. For each browsing
session, we first apply the GEW approach on each
concept for each URL. After applying the GEW to
all concepts, the top β similarities are used to
represent each web page. We select the top β results
because in some cases the concept with the highest
similarity does not give a correct view of a web
page. This could be due to poor or irrelevant
information in a web page, or it could be simply due
to a high level of noise. As a result, we avoid such a
scenario by considering all the top β concepts and
treat them as possible candidates. The context is then
exploited by selecting the common concept that is
associated with different web pages. This common
concept eventually is selected to represent a web
page. If there is no common concept, the concept
with the highest similarity weight is selected. The
full 3C algorithm is described in Algorithm 2.
Input: Similarity results (SR) after applying GEW.
Output: URLs mapped to concepts.
RO= reference ontology.
}, concepts with associated documents.
LOG= log file that contains user's browsing behaviour
LOG ={
} visited web sites
CLU_CON= { TC1, ...,TCn},concepts with total sum of weights
FIN_CLU= final result after applying 3C algorithm
// Select the highest β concepts similarity for each URL.
For each URL
∈SR do
Select top
β concepts;
// find all concepts that appear in different URLs
For each distinct
∈ do
Total_weight= 0;
, Total_weight);
For each 
∈_ do
For each URL
∈SR do
) then
Total_weight += 
End If
, Total_weight);
CLU_CON.sort by Total_weight DESC;
// assign correct concepts to URLs
For each 
∈_ do
For each URL
∈SR do
) then
If FIN_CLU.Does_not_contain(URL
, 
End if
End if
Algorithm 2: Contextual Concept Clustering (3C).
ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence
In order to evaluate the two proposed approaches we
first create a reference ontology using the Open
Directory Project (ODP). Then, we build the
classifier using the TF-IDF classifier. In the next
stage, we create a set of tasks and invite 5 users to
perform these tasks. Finally, we evaluate different
characteristics that impact on the overall
performance of both algorithms, and then we
employ four different mapping approaches to test
and compare them individually. The following
sections describe all these stages in detail.
5.1 Creating a Reference Ontology
For evaluation purposes, we use the ODP concept
hierarchy as a reference ontology (ODP, 2010), and
more specifically the computer category. The
computer directory contains more than 110,000
websites categorized in more than 7000 categories.
In order to train the classifier for each category, all
the websites under each category were fetched.
Furthermore, all the contents of all websites were
extracted and combined in one document. That is,
each category is represented by one document. All
the non-semantical classes (e.g. alphabetical order)
were removed to keep only those classes that are
related to each other semantically. This resulted in a
total of 4116 categories and about 100,000 training
websites whose contents were extracted and
combined in 4116 documents in our reference
ontology. Dimensionality reduction techniques such
as Porter stemming and stop words removal are also
applied to all the 4116 documents. The TF-IDF
classifier (equation 1) is then used to give each term
t in each document d a weight from 0 to 1.
5.2 Collecting Real Usage Data
In order to collect user browsing behaviour, a
Firefox browser is used with a modified add-on
component called Meetimer (Meetimer, 2010). A
SQLite database is used to store all the user’s
sessions. For the purpose of the evaluation process,
35 different concepts from the computer ontology
were selected, and a set of tasks were created. Tasks
took the form of finding a specific piece of
information, or writing a short paragraph. Five users
were invited to perform a total of 90 tasks in a one
month period. For each session throughout the
month, users were asked to select 3 tasks and try to
answer these tasks by browsing and searching for
related web pages. The sequence of the tasks was
not fixed but users were given freedom. After each
session, users were asked to write down what tasks
they performed. These data represent users’ actual
interests which will be matched against the mapped
concepts generated by different mapping
After one month, five log files from five
different users were collected. These five users
together surfed 1,899 web pages. We started
processing the collected data to create processed log
files (P-log files) by fetching all the visited web
pages, and extracting all their contents. We then
applied the GEW and 3C algorithms and compared
the accuracy results against the users’ actual
interests. Next we describe what aspects we have
analyzed and what experiments have been
5.3 Experiments
Three experiments are proposed to analyze different
aspects that impact on the overall performance of
GEW and 3C.
5.3.1 Pruning Non-relevant Concepts
In this experiment, we want to determine a threshold
value (α in GEW algorithm) that could remove non-
relevant concepts to create a more accurate user
profile. For this reason, we apply the GEW
algorithm to: all the retrieved concepts from the
original cosine similarity, top 50, top 20, top 10 and
top 5 results. We use the following accuracy
measure (equation 3) to compute the accuracy.
 =
#     
 
 
Table 1 shows the accuracy percentages for all the
five thresholds after comparing all concepts with the
users’ actual interests. In Table 1, it can be clearly
seen that that the accuracy of applying the GEW
algorithm to all the concepts is relatively low (30%).
While applying GEW to the top 50, 20, 10 and 5
concepts achieved a considerable increase in the
accuracy (71%, 74.90%, 76% and 75.35%
respectively). This shows that applying the GEW to
all concepts could cause inflation in the weight of
the non-relevant concepts. However, applying GEW
on the top 10 results provided the highest accuracy.
This is because the top 10 results could hold the
most important concepts that are likely to be related
to a web page. Based this results, we assign α in the
GEW algorithm to be 10 in the next experiment.
Table 1: Accuracy of all web pages that visited by all
users considering different threshold values.
Threshold Top5 Top10 TOP20 TOP50 All
Precision 75.3% 76.8% 74.9% 71% 30%
In the following experiment, we try to identify
the value of β which is used in the 3C algorithm as a
threshold. In the next experiment, we apply the 3C
algorithm to the top 30, top 20, top 10 and top 5
URLs. Table 2 shows the accuracy percentages for
all the four thresholds after comparing all concepts
with users’ actual interests.
Table 2: Accuracy of using different threshold values for
the 3C algorithm.
Threshold top 5 top 10 top 20 top 30
Accuracy 76.8% 76.1% 75.2% 73.2%
It can be clearly seen from the table 2 that the
accuracy of all the thresholds have achieved close
accuracy results. However, the top 5 threshold
achieved the highest accuracy result, while top 10
and top 20 achieved slightly less accuracy results.
Based on the above results, we assign β in 3C
algorithm to be 5 in all the following experiments.
5.3.2 Rank Ordering
In this experiment, we analyzed the performance of
the 3C algorithm when the concepts are clustered
based on ordering concepts by number of web pages
and by the total similarity weight for each concept.
For both techniques, we calculated the precision for
each user’s profile. Figure 1 shows the ordering
accuracy results for both techniques.
Figure 1: Rank ordering accuracy for each user.
It can be clearly seen that there is a considerable
difference between ordering concepts by number of
web pages and by total similarity weight for each
concept. This could be due to the fact that many
concepts in the log file could share the same super-
concept. As a result, when clustering those concepts
by the number of URLs, the common super-concept
which is likely to be too general is selected.
Consequently, most of the concepts in the user
profile will be too general to represent users’ actual
interests. On the other hand, ordering concepts by
the accumulated weight rather than the number of
URLs, achieved a high average accuracy of about
75.68%. This result demonstrates that clustering and
ordering concepts by the accumulated similarity
weights provides more effective representation of
users’ interests and preferences.
5.3.3 Comparing Mapping Techniques
In this experiment, we aimed at comparing our
mapping techniques (GEW and 3C) to three
different mapping techniques in the literature. The
first technique is the original cosine similarity which
computes the similarity between each URL and all
documents in the ontology. The second technique
which was suggested by Middleton et al. (2004) and
Kim et al. (2007), is adding 50% of each sub-
concept’s weight to its super-concept, and repeats
this process until the root. Finally, the last approach
is the Sub-class Aggregation Scheme that was
proposed by Daoud, Tamine and Boughanem
(2008). For this experiment, each visited web page
for each user was mapped by applying all four
techniques. Figure 2 illustrates the overall accuracy
for each mapping technique for each user.
According to Figure 2, it is clear that the original
cosine similarity and sub-class aggregation schemes
performed poorly for all users (46.17% and 45%
respectively). The main reason that the original
cosine similarity showed the lowest precision is that
the most inaccurate mapped concepts are too
specific and detailed. Similarly, the sub-class
aggregation scheme showed a poor precision of
45%. This is because all the visited web pages were
mapped to only level three classes. The
accumulation behaviour of adding the weights of all
the sub-classes under the level three super-classes
causes inflation on the weights of level three super-
classes. Consequently, no level two classes were
ever mapped to any web page. On the other hand,
adding 50% of each sub-class to its super-class
shows a considerable improvement in the accuracy
average of 60%. This improvement could be
attributed to the fact that if a user is interested in a
User1 User2 User3 User4 User5
Rank ordering accuracy
Ordering by number of URLs
Ordering by total weight
ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence
concept, then he/she also has some interest in its
super-concept. Although this method improved the
overall mapping precision, the percentage which is
added to the super-classes is very high (50%). As a
result, 80% of all the incorrectly mapped web pages
were mapped to level 1 and 2 super-concepts that are
too general to represent user interests. For our
proposed GEW and 3C algorithms, the reported
results were interesting. The overall precision shows
a noteworthy improvement of average of 75%. This
major improvement demonstrates that GEW and 3C
can overcome some of the drawbacks of other
approaches. Furthermore, the GEW and 3C methods
have shown to keep a balance between the general
and the specific interests. Nevertheless, although the
GEW and 3C achieved great results, they have one
limitation. That is, the 3C approach does not take
into the account the semantic relationships between
concepts. In order to improve the performance
further and as part of our future work these
relationships need to be taken into account.
Figure 2: A comparison of 4 different mapping techniques:
OCS: original cosine similarity, 50% from sub-class to its
super-class, sub-class aggregation scheme and GEW &
Web personalization systems enable users to search
for and retrieve information which is tailor-made to
their interests and preferences. However, creating an
accurate user profile unobtrusively and adapting it
dynamically is a complex problem. In this paper, we
presented two novel mapping algorithms (GEW and
3C) that were used to improve the overall accuracy
of the ontological user profile. Our paper revolves
around discovering user interests by mapping visited
web pages to an ontology based on the user
browsing behaviour. Our evaluation results
demonstrate that applying the GEW and 3C mapping
algorithms for modelling user profiles can
effectively improve the overall performance. The
experimental results show that the process of
mapping user interests can be significantly improved
by 28% when utilizing the GEW and 3C algorithms.
As part of further work, we will try to enhance the
mapping process further by exploiting the semantic
relationships between concepts in the ontology.
Azouaou, F. and Desmoulins, C., 2007. Using and
modeling context with ontology in e-learning.
International Workshop on Applications of Semantic
Web Technologies for E-Learning, Dublin, Ireland.
Baeza, R. and Ribeiro, B., 1999. Modern Information
Retrieval. Addison-Wesley.
Challam, V., Gauch, S., and Chandramouli, A., 2007.
Contextual Search Using Ontology-Based User
Profiles, Proceedings of RIAO 2007,Pittsburgh, USA.
Chen, C., Chen, M., Sun, Y., 2002. PVA: A Self-Adaptive
Personal View Agent System. Journal of Intelligent
Information Systems, vol. 18, no. 2, pp. 173–194.
Daoud, M., Tamine, L., Boughanem, M., 2008. Learning
user interests for session-based personalized search. In
ACM Information Interaction in context (IIiX).
Felden, C. and Linden, M., 2007. Ontology-Based User
Profiling. Lecture Notes in Computer Science. vol.
4439, pp. 314-327.
Gorgoglione, M., Palmisano, C. and Tuzhilin, A., 2006.
Personalization in Context: Does Context Matter
When Building Personalized Customer Models?
ICDM 2006. pp. 222-230.
Kim H., Chan P.,2003. Learning Implicit User Interest
Hierarchy for Context in Personalization. In
Proceedings of the 2003 International Conference on
Intelligent user interfaces 2003, Miami, Florida.
Kim, J., Kim, J., Kim, C., 2007. Ontology-Based User
Preference Modeling for Enhancing Interoperability in
Personalized Services. Universal Access in HCI, Part
III, HCII 2007, LNCS 4556. pp. 903–912.
Lin, R., Kraus, S. and Tew, J., 2004. Osgs - a personalized
online store for e-commerce environments.
Information Retrieval journal, vol.7(3–4).
Liu, F., Yu, C., Meng, W., 2002. Personalized web search
by mapping user queries to categories. In Proceedings
of CIKM. Virginia, vol.02, pp.558-565.
Meetimer, 2010.
US/firefox/addon/5168/, April 2010.
Middleton, S., Shadbolt, N., Roure, D., 2004. Ontological
user profiling in recommender systems. ACM Trans.
Open Directory Project, 2010., April
Pan, X., Wang, Z. and Gu, X., 2007, Context-Based
Adaptive Personalized Web Search for Improving
user1 user2 user3 user4 user5
Comparing mapping techniques
OCS 50% to super-clacc SCAS GEW and 3C
Information Retrieval Effectiveness, International
Conference on Wireless Communications, Networking
and Mobile Computing, 2007. pp.5427-5430.
Pignotti E., Edwards, P. and Grimnes, G., 2004. Context-
Aware Personalised Service Delivery. European
Conference on Artificial Intelligence, ECAI 2004.
Porter, M., 1980. An algorithm for suffix stripping.
Program 14,3,pp.130-137.
Sieg, A., Mobasher, B. and Burke, R., 2007. Representing
context in web search with ontological user profiles. In
Proceedings of the Sixth International and
Interdisciplinary Conference on Modeling and Using
Context. Roskilde, Denmark.
Trajkova, J., Gauch, S., 2004. Improving ontology-based
user profiles. In Proceedings of RIAO.
Van, C., Robertson, S., Porter, M., 1980. New models in
probabilistic information retrieval. British Library
Research and Development Report, no. 5587. London.
ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence