AN ONTOLOGY-BASED APPROACH TO PROVIDE

PERSONALIZED RECOMMENDATIONS USING A

STOCHASTIC ALGORITHM

Romain Picot-Clémente, Christophe Cruz and Christophe Nicolle

LE2I Laboratory, University of Bourgogne, Dijon, France

Keywords: Recommender systems, Information filtering, Semantic web, Adaptive hypermedia systems, User

modelling, Stochastic processes.

Abstract: The use of personalized recommender systems to assist users in the selection of products is becoming more

and more popular and wide-spread. The purpose of a recommender system is to provide the most suitable

items from an knowledge base, according the user knowledge, tastes, interests, ... These items are generally

proposed as ordered lists. In this article, we propose to combine works from adaptive hypermedia systems,

semantic web and combinatory to create a new kind of recommender systems suggesting combinations of

items corresponding to the user.

1 INTRODUCTION

In the first years after its creation, Web was the same

for everybody. Websites was presenting the same

information and the same links to each visitor,

without taking into account their goals and their

knowledge.

Given the important growth of the number of

available information, the diversity of users and the

complexity of Web applications, researchers started

to question the "one-size-fits-all" approach.

Therefore researches have emerged to answer these

problems, to propose a Web where the appearance,

the behaviour, the resources would be ideally

adapted to each individual. This Web would offer a

different experience to each user. This Web is called

the Adaptive Web.

A first step to reach the Adaptive Web is to

reason on local adaptation problems.

This adaptation to users is currently widely

found into the so-called domain of recommender

systems (Tintarev and Masthoff, 2007) (Mcsherry,

2005). These systems have the role to provide

recommendations to users. This domain has become

important since the mid-year 90. It is yet a big

domain of research due to its richness in terms of

problem to solve, and given the great number of

possible applications to deal with information

overload and provide adaptive recommendations.

Recommender systems allow companies to filter

information, and then to recommend products to

their customers, according their preferences.

Recommending products or services can consolidate

relations between the seller and the customer, and

thus, enhance the benefits (Zhang and Jiao, 2007).

Many systems developed as TrustWalker (Jamali

and Ester, 2009) emphasize the current interest of

recommendation.

Content-based (CB) and collaborative filtering

(CF) methods are two of the main approaches used

to form recommendations. Hybrid techniques

integrating these two different approaches have also

been proposed. The CB method has been developed

basing on the textual filtering model described in

(Oard, 1997). Generally, in CB systems, the user

profile is inferred automatically from documents'

content that the user has seen and rated. The profiles

and domain documents are then used as input of a

classification algorithm. The documents which are

similar (in content) to the user profile are considered

interesting and are recommended to the user.

CF systems (Goldberg, 1992) are an alternative

to CB systems. The basic idea is to go beyond the

experience of an individual user profile and instead

to use the experiences of a population or community

of users. These systems are designed with the

assumption that a good way to find interesting

content is to find people with similar tastes and to

659

Picot-Clémente R., Cruz C. and Nicolle C..

AN ONTOLOGY-BASED APPROACH TO PROVIDE PERSONALIZED RECOMMENDATIONS USING A STOCHASTIC ALGORITHM.

DOI: 10.5220/0003478506590665

In Proceedings of the 7th International Conference on Web Information Systems and Technologies (SWAT-2011), pages 659-665

ISBN: 978-989-8425-51-5

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

propose items they like. Typically, each user is

associated to a set of nearest-neighbour users,

comparing profiles’ information. With this method,

objects are recommended basing on similarities of

users rather than the similarities of objects.

Both CF and CB systems have strengths and

weaknesses. In CF systems, the main problem is that

the new objects with no rate cannot be

recommended. CB systems suffer from deficiencies

in the way of selecting items for recommendation.

Indeed, the objects are recommended if the user has

seen and liked similar objects in the past.

Consequently, a variety of hybrid systems have

recently been developed:

 Some use other users’ ratings as additional

features in a CB system.

 Some use CB methods for the creation of bots

producing additional data for “pseudo-users”.

These data are combined with real users’ data

using CF methods.

 Others use CB predictions to “fill out” the

probable user-items’ ratings in order to allow

CF techniques to produce more accurate

recommendations.

In this paper, we propose a content-based

recommender system. Its architecture is inspired

from adaptive hypermedia researches. Unlike most

recommender systems which provide a list of items

to users, we suggest to generate combinations of

items corresponding to them. A first part will

explain what an adaptive hypermedia system is.

Then, our proposition is presented. The final part

shows a application to tourism domain.

2 ADAPTIVE HYPERMEDIA

SYSTEMS

The research domain of adaptive hypermedia has

been very prolific these 10 last years. Some systems

(Cristea and Mooij, 2003) (Henze, 2000) have been

developed, giving principally solutions for e-

Learning which is considered as the first application

domain. Each system brings its own architecture and

methods. Moreover, few attempts have been made to

define reference models (De Bra et al., 1999)

(Hendrix and Cristea, 2008) but without success

because of being not enough generic to take account

of the new trends and innovations. Nevertheless,

most of the systems and models are based on a set of

layers, also called models, which separate clearly the

different tasks. Then, we can see that there are at

least three models in common, necessary and

sufficient to achieve adaptive hypermedia systems

according Brusilovsky (Brusilovsky, 2001). It needs

to primarily be a hypermedia system based on a

domain. The domain model is a representation of the

knowledge on a given subject the creator wants to

deliver. It describes how the domain is organized

and interconnected. The second model is called a

user model which is a representation of the user

within the system. It models all user information

which may require the system to provide an

adaptation. The last model is the adaptation model.

It performs all the adaptive algorithms based on

other models to provide an adaptation to the user.

Beyond the use of domain, user and adaptation

models, the trend is to use additional models like

presentation, goals, context or other models. This

allows to better identify the different performed

tasks and to facilitate the construction of adaptive

hypermedia systems. Nevertheless, there is no

generic model integrating them for the moment.

Methods to model domain/user (adaptation

principles are also described):

 Keywords vectors: A popular method to

represent domain and user model used by a lot

of systems is the keywords vectors space

representation (Liberman, 1995) (Kamba et

al., 2007). This method considers that each

document and user profile is described by a

set of weighted keywords vectors. At the

adaptation model, the weights are used to

calculate the similarity degree between two

vectors and then to propose relevant document

to the user. The keywords representation is

popular because of its simplicity and its

efficiency. Nevertheless, the main drawback is

that a lot of information is lost during the

representation phase.

 Semantic networks: In semantic networks,

each node represents a concept. Minio and

Tasso (Minio and Tasso, 1996) present a

semantic networks based approach where each

node contains a particular word of a corpus

and arcs are created following the co-

occurrences of the words from connected

nodes into the documents. Each domain

document is represented like that. In simple

systems using only one semantic network to

model the user, each node contains only one

keyword. The keywords are extracted from

pages which the user gives its taste. Then,

they are treated to keep only the most relevant

ones and are weighted in order to remove

those with a weight lesser than a predefined

WEBIST 2011 - 7th International Conference on Web Information Systems and Technologies

660

threshold. The selected keywords are then

added to the semantic network where each

node represents a keyword and each arc their

co-occurrence into the documents. With this

method, it is possible to evaluate the relevance

of a document compared to the user profile.

Indeed, it suffices to construct a semantic

network of a document and compare it to the

semantic network of the user to classify it to

interesting, uninteresting or indifferent

documents.

 Ontology: This approach is similar to the

semantic network approach in the sense that

both are represented using nodes and relations

between nodes. Nevertheless, in concepts

based profiles, nodes represent abstract

subjects and not word or set of words.

Moreover, links are not only co-occurrence

relations between words, they have several

significations. The use of ontology can keep a

maximum of information. In QuickStep

(Middleton at al., 2002), the ontology is used

for the research domain and has been created

by domain experts. The ontology concepts are

represented as vectors of article examples. The

users’ papers from their publication list are

modelled as characteristic’ vectors and are

linked to concepts using the nearest neighbour

algorithm. These concepts are then used to

form a user profile. Each concept is weighted

by the number of papers linked to it.

Recommendations are then made from the

correlations between the current interests of

the user to topics and papers that are related to

these topics. In (Cantador and Castells, 2006)

and (Sieg et al., 2007), a predefined ontology

is used to model the domain. User profiles are

represented with a set of weighted concepts

where weight represents the user’s interest for

a concept. Its interests are determined by

analyzing its behaviour.

Three types of adaptation have been highlighted

in the researches on adaptive hypermedia systems:

content, navigation and presentation adaptation. The

content adaptation consists in hiding/showing or

highlighting some information. The adaptation

model makes the decision of which content has to be

adapted and how to display it. The navigation

adaptation consists in modifying the link structure

suggesting links or forcing the user to follow a

destination. There is URLs’ adaptation or

destinations adaptation. In the first, the adaptation

model provides destination links to the presentation

model; these links are displayed at the page

generation. Whereas, in the second one, the

adaptation model provides links without fixed

destination to the presentation model; the destination

is decided by the adaptation model when the link is

accessed by the user. The presentation adaptation

consists in insisting (or not) on the content parts or

on the links. It consists also in adapting the

preferences setting to the device or the page. The

adaptation model process makes the decision of

which content or links to insist in following the

presentation context.

Even if recommender systems are often

differentiated from adaptive hypermedia systems, a

lot of similarity between these two types of systems

can be highlighted. Indeed, the recommender

systems provide recommendations using different

algorithms, as it is done in the adaptation model.

Moreover, we can see that they model also users’

tastes and domain’s items, as it is done in adaptive

hypermedia systems with the user model and domain

model. Nevertheless, recommender systems perform

only adaptation of the content whereas adaptive

hypermedia systems realize two more adaptation

types. Following these observations, a

recommendation system appears to be a constrained

adaptive hypermedia system. Thus, it seems clear

that recommender systems can be defined as a

subset of adaptive hypermedia systems, whatever its

type (CB or CF).

The use of an adaptive hypermedia architecture

for the creation of recommender systems is

interesting because we can clearly define the tasks

associated with each part of the application, and it

gives the opportunity to evolve the system adding

modules and/or other types of adaptation without

difficult modifications of parts already implemented.

For instance, a CB recommender system could be

improved with features of CF systems, adding a

group model where clusters of users can be defined.

For the creation of our CB recommender system,

we base on adaptive hypermedia architecture.

Beyond the use of the three main ones (domain, user

and adaptation model), a goal model has been added.

It allows the modelling of users’ goals. A description

of the proposition is explained in the following part.

3 PROPOSITION

This section describes the architecture of the

proposed recommender system based on an adaptive

hypermedia architecture. This system aims to

generate a combination of necessary items for the

purpose of the application according the items types

AN ONTOLOGY-BASED APPROACH TO PROVIDE PERSONALIZED RECOMMENDATIONS USING A

STOCHASTIC ALGORITHM

661

Figure 1: The components of the architecture.

and constraints (semantics, localization, ...). An item

is a piece of information provided by the

information system transformed into the

recommender system. The architecture of the system

is composed of four components: The domain model

component, the goal model component, the user

model component, the adaptation model component

(Figure. 1).

3.1 The Domain Model Component

To fully express the knowledge of a domain, the

domain model component uses an OWL ontology.

An item will be an individual of this domain

ontology which is generated from data provided by

data sources like relational databases, XML

documents, Excel files, etc. Actually, the ontology is

enriched by the schema of the data sources and

populated by the data of the sources. This

construction is carried out using expert’s knowledge.

Three steps have been identified to achieve this goal.

The first step consists in using the database as a

starting point for the basis of the domain ontology.

The second step consists in enriching this basic

ontology by domain experts adding new axioms

usually not present in the data source. Usually, this

process comprises the addition of semantic links

between concepts which are often unqualified in the

relational schema. Moreover, the hierarchical

structure of concepts is refined with the help of

professionals. The last step consists in populating

the ontology with instances selected from the data

source.

The ontology creation process is partly manual

and automatic. The first step and the second step

require the investment of domain experts. The third

step is automatic and, fortunately, it is the only one

that is brought to be repeated regularly to update the

data of the ontology. The structural part of the

ontology which is set up by the first two steps will

not often be changed.

3.2 The Goal Model Component

Our objective is to use the domain ontology within a

specific user-centered application. In this context, an

adjustment of the ontology must be made to adapt it

to the usage and, especially, to the user modeling.

Hence, the domain ontology is enriched with goal

concepts. These concepts define the user possible

objectives into the system.

A goal concept is composed of first order logic

rules. A rule makes a link between individuals of the

ontology and the related goal. Each rule has a weight

representing the importance of it within a goal

concept.

WEBIST 2011 - 7th International Conference on Web Information Systems and Technologies

662

For example, we could have the goal

"culturalActivity " defined by the weighted rules:

- (Cinema(x) ^ haveCulturalAspect(x,y) ->

culturalActivity (x), 5)

- (Theater(x) ^ haveCulturalAspect(x,y) ->

culturalActivity (x), 9)

- (Museum(x) ^ haveCulturalAspect(x,y) ->

culturalActivity (x), 10)

The set of the goal concepts is called the goal

model.

3.3 The User Model Component

From there, we have an ontology oriented to a

particular application by a set of goal concepts. We

have now to model the user within the system in

order to offer an adapted environment. The user is

modeled in two distinct parts. The first one concerns

the characteristics that are independent of the

domain, also called the static part. This may be the

age, size, gender, etc. They are simply modeled

using <attribute, value> vectors. The second part

represents the user features directly related to the

domain. It consists of goal concepts, representing the

goals of the user with weights which allow the

position of a discriminating index in terms of

preferences. This part is said dynamic and is an

overlay on the domain model. Together, these two

parts represent the user model and the instantiation

of this model represents a profile of a user.

For instance, a user profile can be: User1 (<age,

20>, <gender, male>, …) (<physicalActivity,5>,

<culturalActivity,1>, …)

3.4 The Adaption Model Component

This component is responsible to provide a

combination of items regarding the domain

ontology. This set of item is adapted to the user

profile which is an instance of the user model. In

order to enclose the solution to provide, a pattern of

combination is defined. The pattern constrains the

elements to be returned to the user by filtering to the

type, the number, the item order or the

geolocalisation for example.

Using the goals weights into the user profile and the

weighted connections between the individuals and

goal concepts, the weights are combined linearly to

be propagated to the individuals.

For instance, if a goal is weighted of 3 in a

profile and is constituted of two rules with the

respective weights 5 and 1, then the individuals

linked to the goal by these rules receive the

respective weights

)53(15

and

)13(3 ×

which can

be cumulated if individuals are associated to the two

goals at once.

This weight propagation into the ontology is not

done with the whole ontology. Indeed, it is not

necessary to weight individuals which cannot be

involved in the final solution. Only individuals that

instantiate a concept from the domain model

specified into the pattern of combination are

weighted. In other words, the individuals which the

type (the concept) is not reflected in the pattern are

not weighted. Thus, we have a subset of weighted

individuals from the domain ontology. We now wish

to find a "good" combination of these individuals

depending on the pattern. The notion of quality of a

combination is defined by one or more relevance

functions that will take into account the individual

weights and other attributes (geographic coordinates,

volume, etc.).

For example, we could have this pattern <

(Activity, Activity, Activity, Activity), 1km > which

constrains the combination to have 4 activities in a

neighborhood of 1km.

The purpose consists then in optimizing the

relevance function, in order to find the optimal

solution. Searching an optimal solution means

browsing the set of possible solutions and returns the

best one. However, the number of possible solutions

grows very rapidly according to the number of

individuals, and it becomes quickly impossible to

find the best solution in a reasonable time. Thus, to

resolve this problem, a stochastic algorithm is used.

This latter takes as input all the weighted individuals

and the pattern of combination to provide. This

algorithm will return “good” items approaching the

best one. However, the results will depend on the

type of the used stochastic algorithm which optimize

the time for real-time application. Finally, the user is

offered a combination of items supposed to

correspond to their profile. The adaptation model is

composed of all these processes.

4 TOURISM APPLICATION

This modeling is being applied to the tourism

domain in the region of Côte-d’Or in France for the

company Côte-d’Or Tourisme. The aim is to create a

tourism application that should provide a

combination of tourism products from Côte-d’Or

according to a user profile. At the beginning, a

domain ontology has been created with all the

concepts and the individuals related to the

application domain. This ontology was supplied

from a database composed of more than 4000

AN ONTOLOGY-BASED APPROACH TO PROVIDE PERSONALIZED RECOMMENDATIONS USING A

STOCHASTIC ALGORITHM

663

tourism products. Then, a goal model has been

defined using goal concepts like “Week end”,

“Going out with friends”, “with a baby”, etc. This

knowledge was generated from the specialists of the

domain represented by people working for the

company Côte-d’Or Tourisme.

An empirical pattern is defined to determine

what kind of combination the adaptation model has

to return. The relevance function giving the

relevance of a combination is based on the interest

weights and the coordinates of the tourism products,

because it is not relevant to propose an activity in

the morning and a restaurant for lunch with a

distance of more than 50 kilometers. The traveling

time required to reach the restaurant after the

activity ending is inappropriate.

A variance threshold needs to be set in order to

define the maximum preferred variance between the

individuals coordinates. This variance characterized

the value dispersion regarding the average, in this

case the threshold. Subsets in this pattern are

possible. For instance, we can define a pattern like

“Accommodation, Restaurant1, Activity,

Restaurant2” in which “Hotel and Restaurant1” are

the first subset, and “Activity and Restaurant2” the

second subset. In addition, a variance threshold is

defined for each one. Thereby, the system can use

more complex patterns for the combinations.

The variance of a combination is defined as

follow:

11 1

00 0

11 1

var( )

NN N

ix ix iy iy

ii i

CCCCC

NN N

−− −

== =

⎛⎞

⎛⎞⎛⎞

=−+−

⎜⎟

⎜⎟⎜⎟

⎜⎟

⎝⎠⎝⎠

⎝⎠

∑∑ ∑

(1)

where C is a combination, N the number of

elements into the combination, C

the i

element of

the combination, and C

and C

the x and y

coordinates of the i

element.

The weight of a combination is defined as

follow:

∑

−

)(

iweight

Cweight

(2)

where C

iweight

is the weight of the i

element.

Using the variance and the weight function, the

relevance of a combination is:

var( )

1var()

()

Energy C E E

weight C Threshold Threshold

−

⎛⎞

⎜⎟

=× +

⎜⎟

⎝⎠

∑

(3)

where E(X) is the integer part of X, Threshold

is the variance threshold of the geographic

coordinates for the combination C,

Threshold

the variance threshold of the geographic coordinates

for the j

subset of C, and L the number avec

subsets.

Then, a stochastic algorithm is performed. We

use a simulated annealing algorithm (Kirkpatrick et

al., 1983) which try to optimise the relevance

function, so that it can provide a combination of

tourism products matching the user interests and

with close coordinates.

5 BENCHMARK

Some tests of the algorithm for the generation of

combinations have been done on a set of four

thousand tourism products. In these tests, the

average time required for the generation of a

combination of six products was around 3 seconds.

But, this time depends on the different parameters

(temperature decrease rate, the number of iteration

per level of temperature) necessary to perform the

simulated annealing algorithm. For example, the

faster is the temperature decrease and the lower is

the number of iterations, the faster is the generation,

but the worse is the resulting combination. In any

case, this time is better than the time required to find

the best combination by browsing all the

possibilities. For example, in our test, finding the

best combination needed around 3 hours against 3

seconds using the simulated annealing algorithm.

These times are only given to have orders of

magnitude, more tests need to be performed to have

exacts results and to prove the interest of our

proposition. Nevertheless, given these few results,

the algorithm seems to give a relevant solution

according a predefined relevance function with a

lesser cost (in terms of time) than calculating the

optimal solution.

6 CONCLUSIONS

This paper presents a new content based

recommender system. The idea consists in taking

advantages of the semantic Web technologies, the

properties of adaptive hypermedia systems, and in

combining them with combinatory algorithms in

order to provide recommendations, adapted

combinations of items. The simulated annealing

algorithm is used in order to solve the problem of

the polynomial time search required to generate a

combination of tourism products. It gives a solution

which approaches the best solution in a short time.

The few results seem to be good considering the

time required to obtain them and comparing to the

best solutions. Nevertheless, for the future, we need

WEBIST 2011 - 7th International Conference on Web Information Systems and Technologies

664

to make more tests and benchmarks to quantify more

precisely the relevance of our system. Moreover, we

could improve the quality of the propositions by

taking into account some group of users as it is done

in collaborative filtering recommender systems. It is

possible by adding a group model into the

architecture. Thus, the recommender system would

become a hybrid recommender system.

ACKNOWLEDGEMENTS

This research is done in collaboration with Côte-

d’Or Tourisme, a company which aims to promote

tourism in its region.

REFERENCES

Brusilovsky, P. 2001. Adaptive hypermedia, User

modeling and user-adapted interaction, 2001 –

Springer.

Cantador, I. and Castells, P. 2006. A multilayered

ontology-based user profiles and semantic social

networks for recommender systems. In 2

International Workshop on Web Personalisation

Recommender Systems and Intelligent User Interfaces

in Conjunction with 7th International Conference in

Adaptive Hypermedia.

Sieg, A., Mobasher, B. and Burke, R. 2007. Leaning

Ontology-Based User Profiles: A Semantic Approach

to Personalized Web Search. In IEEE Intelligent

Informatics Bulletin, vol.8, no. 1.

Cristea, A. and De Mooij, A. 2003. LAOS: Layered

WWW AHS authoring model and their corresponding

algebraic operators, Proceedings of World Wide Web

International Conference, New York, NY: ACM.

De Bra, P., Houben, G-J. and Wu, H. 1999. AHAM: A

dexter-based reference model for adaptive

hypermedia, in Hypertext’99: Proceedings of the 10th

ACMConference on Hypertext and Hypermedia:

Returning to our Diverse Roots, New York, pp.147-

156.

Goldberg, D., Nichols, D., Oki, B. M. and Terry, D. 1992.

Using collaborative filtering to weave an information

tapestry. Communications of the Association for

Computing Machinery, 35(12):61–70.

Hendrix, M. and Cristea, 2008. A. Meta-levels of

adaptation in education, Proceedings of 11th IASTED

International Conference on Computers and Advanced

Technology in Education, V. Uskov (Ed.), IASTED.

Henze, N. 2000. Adaptive hyperbooks: Adaptation for

project-based learning resources. PhD Dissertation,

University of Hannover.

Jamali, M. and Ester, M. 2009. TrustWalker : a Random

Walk Model for Combining Trust-Based and Item-

Based Recommendation. In KDD ’09 : Proceedings of

the 15th ACM SIGKDD international conference on

Knowledge discovery and data mining, New York,

NY, USA, pp. 397–406. ACM.

Kamba, T., Sakagami, H. and Koseki, Y. 2007.

Antagonomy: A Personalized Newspaper on the

World Wide Web, Int’l J. Human-Computer Studies.

Kirkpatrick, S., Gelatt C. D. and Vecchi, M. P. 1983

Optimization by simulated annealing, Science 220

(4598), 671-680.

Liberman, H. 1995. Letizia: An Agent that Assists Web

Browsing. In Proceedings of the Fourteenth

International Joint Conference on Artificial

Intelligence (IJCAI-95), pp 924--929, Morgan

Kaufmann publishers Inc, Montreal, Canada.

Middleton, S., Alani, H., Shadbolt, N. and De Roure, D.

2002. Exploiting synergy between ontologies and

recommender systems. In Proceedings of the WWW

international workshop on the semantic web, v. 55 of

CEUR workshop proceedings, Maui, HW, USA.

Minio, M. and Tasso, C. 1996. User Modeling for

Information Filtering on INTERNET Services:

Exploiting an Extended Version of the UMT Shell. In

UM96 Workshop on User Modeling for Information

Filtering on the WWW; Kailua-Kona, Hawaii, January

2-5.

Mcsherry, D. 2005. Explanation in recommender systems.

Artificial Intelligence Review, 24(2):179 – 197, 2005.

Oard, D. 1997. The State of the Art in Text Filtering User

Modeling and User Adapted Interaction, 7(3)141-178.

Tintarev, N. and Masthoff, J. 2007. A survey of

explanations in recommender systems,

ICDE'07

Workshop on Recommender System.

Zhang, Y. Y. and Jiao, J. X. 2007. An associative

classification-based recommendation system for

personalization in B2C e-commerce application,

Expert Systems with Applications 33 (1) 357–367.

AN ONTOLOGY-BASED APPROACH TO PROVIDE PERSONALIZED RECOMMENDATIONS USING A

STOCHASTIC ALGORITHM

665