A CASE STUDY ON DOMAIN ANALYSIS OF SEMANTIC WEB

MULTI-AGENT RECOMMENDER SYSTEMS

Roberval Mariano, Rosario Girardi, Adriana Leite, Lucas Drumond and Djefferson Maranhão

Federal University of Maranhão, Av. dos Portugueses, Campus do Bacanga, São Luís–MA, Brazil

Keywords: Recommender Systems, Semantic Web, Information Filtering, Domain Engineering, Multi-agent Systems.

Abstract: The huge amount of data available on the Web and its dynamic nature is the source of an increasing

demand of information filtering applications such as recommender systems. The lack of semantic structure

of Web data is a barrier for improving the effectiveness of this kind of applications. This paper introduces

ONTOSERS-DM, a domain model that specifies the common and variable requirements of Recommender

Systems based on the ontology technology of the Semantic Web, using three information filtering

approaches: content-based, collaborative and hybrid filtering. ONTOSERS-DM was modeled under the

guidelines of MADEM, a methodology for Multi-Agent Domain Engineering, using the ONTOMADEM

tool.

1 INTRODUCTION

Recommender systems suggest items of user´s

interests using information filtering techniques

(Adomavicius and Tuzhilin, 2005), which can be

classified in:

• Content Based Filtering (CBF), where an item

is recommended in terms of its similarity with

other items that the user evaluated positively in

the past.

• Collaborative Filtering (CF), where an item is

recommended according to the last preferences

of other users with similar interests.

• Hybrid Filtering (HF), where the two previous

filtering approaches are combined.

The effectiveness of recommender systems is

limited in some environments, such as the Web, due

to lack of semantic mark up of the information

contained in this environment. An approach to this

problem is the Semantic Web, where the

information is represented in knowledge

representation structures like ontologies.

This work describes the domain modeling task of

the MADEM (“Multi-agent Domain Engineering

Methodology”) methodology (Girardi and Marinho,

2007) and introduces ONTOSERS-DM, a domain

model that specifies the common and variable

requirements of a software family of multi-agent

recommender systems based on the Semantic Web

technologies.

The paper is organized as follows. In section 2,

an overview of the domain analysis phase of the

MADEM methodology is introduced. Section 3

describes the development of ONTOSERS-DM.

Section 4 analyses some related work on similarly

developed domain models and on the application of

the Semantic Web technologies in recommender

system development. Finally, section 5 concludes

this paper with a discussion on further work being

conducted.

2 AN OVERVIEW OF THE

MADEM DOMAIN ANALYSIS

PHASE

The modeling concepts of the MADEM

methodology as well as its products are represented

as instances of the ONTOMADEM ontology, a

knowledge-based tool that supports the modeling

tasks of MADEM (Girardi and Leite).

The modeling concepts, tasks and products of

MADEM are based on techniques for Domain

Engineering and Multi-agent System Development.

MADEM is divided in the domain analysis, domain

160

Mariano R., Girardi R., Leite A., Drumond L. and Maranhão D. (2008).

A CASE STUDY ON DOMAIN ANALYSIS OF SEMANTIC WEB MULTI-AGENT RECOMMENDER SYSTEMS.

In Proceedings of the Third International Conference on Software and Data Technologies - ISDM/ABF, pages 160-167

DOI: 10.5220/0001885801600167

 SciTePress

Figure 1: Some relationships between concepts and its instances in ONTOSERS-DM.

design, domain implementation and pattern

extraction and representation phases. Each phase has

its respective set of tasks and generated products.

The domain analysis phase consists of the concept

modeling, goal modeling, role modeling, role

interactions modeling and user interface prototyping

tasks (Table 1). The concept modeling task

represents a brainstorming of main domain concepts.

The goal modeling task represents the general goal,

external entities and specific goals. The role

modeling task represents the detailing of each

specific goal, with indication of the associated roles,

knowledge, pre-conditions, post-conditions and

external entities. The role interactions modeling task

defines how the roles cooperate to achieve their

goals. The user interface prototyping task aims at

constructing a prototype of the requirements

expressed in the Goal, Role and Role interactions

models. Variability modeling is carried out in

parallel with goal, role and role interactions

modeling to determine the common and variable

parts of an application family. This is done by

identifying the “Variation Points” and its

correspondent “Variants”. A variation point is the

representation of a concept subjected to variation. A

variant represents the alternative or optional

variations of such concept. For example, in Figure 1

the “Variation Point 1” indicates that the specific

goal “Model User” has three alternative variants,

being each variant an instance of the “responsibility”

concept. A general goal is reached from the

execution of a set of specific goals (relationship

“reached from” of Figure 1). Specific goals are

reached from the execution of responsibilities

(relationship “achieves”). A specific goal can have

as variation points responsibilities that are selected

according to the characteristics of each particular

application.

In the example of Figure 1, the “Model users”

specific goal has a variation point with alternative

groups of responsibilities for user profile

acquisition, being possible choosing between three

variants: “Implicit Profile Acquisition”, “Explicit

Profile Acquisition” or both. Each role is in charge

of one responsibility (relationship “performed by”).

Pre-conditions and post-conditions must be satisfied

(relationship “is satisfied”). The participant entities

can be external, that interact with the system, or

internal who play roles (relationship “participates”).

3 DOMAIN ANALYSIS OF

ONTOSERS-DM

This section describes the development of the tasks

and their respective products, of the domain analysis

phase of the MADEM methodology for the

construction of the domain model ONTOSERS-DM.

A CASE STUDY ON DOMAIN ANALYSIS OF SEMANTIC WEB MULTI-AGENT RECOMMENDER SYSTEMS

161

Table 1: Resources, tasks and products of the domain analysis phase of the MADEM methodology.

Figure 2: Concept Model of the ONTOSERS-DM.

3.1 Concept Modeling

The Concept Model of the ONTOSERS-DM is

illustrated in Figure 2. The main difference between

traditional recommender systems and the “Semantic

Web-based Recommender Systems” are that the last

one uses “Information Item Models” that are a kind

of “Semantic Web Resources”. This way, they take

advantage of the Semantic Web technologies to

represent “Information Item Models” and “User

Models” as instances of a “Domain Ontology”. A

“Personalized Recommendation” is constructed

based on an “Information Item Model” and “User

Models”. In the case of Content-Based Filtering, the

recommendation is based on an “Information Item

Model”. In the case of Collaborative Filtering, the

recommendation is based on a “Model of Group of

Users with Similar Profiles”. A “Personalized

Recommendation” satisfies a “User Profile”. Such

recommendation is composed by “Filtered Items”

and is delivered directly to the user.

3.2 Goal Modeling

Figure 3 represents the goal model of ONTOSERS-

DM. The “Provide Recommendations using

Semantic Web Technology” general goal is reached

through the “Model Users”, “Filter Information” and

“Deliver Recommendations” specific goals. Figure 4

shows the variation points of the specific goals

defining the variability of the responsibilities needed

to reach this goal. In order to achieve the “Filter

Information” specific goal, it is necessary to perform

the “Ontology Instance User Model Creation and

Update” responsibility, which also contributes to

reach the “Model Users” specific goal. Besides that,

the “Grouping of user models”, “Information Items

based on Ontology Instance Representation” and

“Similarity Analysis” responsibilities are needed.

The “Grouping of Users Models” responsibility

allows identifying groups of users with similar

interests.

The “Information Items based on Ontology

Instance Representation” allows representing

information items in a structure that can be

processed by software agents. The “Similarity

ICSOFT 2008 - International Conference on Software and Data Technologies

162

Figure 3: Goal Model of ONTOSERS-DM.

Analysis” responsibility looks for determining

the relevance of a given item for a user. The “Filter

Information” specific goal has a variation point that

has as variant alternatives: the “Grouping of users

models” responsibility, required in systems that use

CF; and the “Information Items based on Ontology

Instance Representation” responsibility in the ones

using CBF. To reach the “Deliver

Recommendations” specific goal is necessary to

perform the “Similarity Analysis” responsibility.

The “Deliver Recommendations” specific goal is

also reached through the execution of the

“Production of Personalized recommendations” and

“Deliver Personalized Recommendations”

responsibilities. The “Model Users” specific goal

has a variation point with groups of responsibilities

for user profile acquisition, being possible to choose

between three alternative variants: “Implicit Profile

Acquisition”, “Explicit Profile Acquisition” or both.

The last responsibility, “Ontology Instance User

Model Creation and Update” is fixed, i.e. it is

required in all the applications of the family. The

“Deliver Recommendations” specific goal does not

have variation points, therefore the “Similarity

Analysis”, “Personalized Recommendations

Production” and “Delivery of Personalized

Recommendations” responsibilities are required in

all the applications of the family, then belonging to

the fixed part of the goal model.

3.3 Role Modeling

In the Role Modeling task, each one of the

responsibilities identified in the Goal Model is

assigned to an internal entity playing a role. This is

expressed in a Role Model. Each role responsibility

requires the usage and the production of certain

knowledge and the fulfillment of pre-conditions and

post-conditions. Figure 5 shows the role model

related to the “Model Users” specific goal. The

“User Monitor” role plays the “Implicit Profile

Acquisition” responsibility. This responsibility has

the pre-condition “User is connected”. The “Explicit

Profile Acquisition” responsibility, associated to the

“Input Interface” role, can be performed if the “User

filled a form” pre-condition is satisfied. This

responsibility produces a “User Profile” and has a

“User Profile Acquired” post-condition. To execute

the “Ontology Instance User Model Creation and

Update” responsibility, the “User Modeler” role uses

a “User Profile” and produces an “Ontology based

User Model” thus having the “User models are

created” post-condition satisfied.

Figure 6 shows the Role Model related to the “Filter

Information” specific goal. The “Grouping of User

Models” responsibility is executed by the “Miner”

role. This responsibility uses a “User Model” and

produces “Groups of Users Models” thus satisfying

the “Groups Available” post-condition. When a

“New Item of Information is available” pre-

condition is satisfied, the “Information Item based

on Ontology Instance Representation” responsibility

is exercised using a “New Information Item” and

producing an “Information Item Model”, thus

satisfying the “Information Item is represented”

post-condition.

The “Similarity Analysis” responsibility is executed

by the “Filter” role for determining how much an

information item can satisfy the information needs

of a user, represented in a User Model.

A CASE STUDY ON DOMAIN ANALYSIS OF SEMANTIC WEB MULTI-AGENT RECOMMENDER SYSTEMS

163

Figure 4: Variation Point of the Specific Goal Model Users.

Figure 5: Role Model of the “Model Users” Specific Goal.

This responsibility uses a “Groups of Users Model”

and/or an “Information Item Model” and produces a

set of “Filtered items”. The “Deliver

Recommendations” specific goal is reached through

the performance of the “Similarity Analysis”,

“Personalized Recommendations Production” and

“Delivery of Personalized Recommendation”

responsibilities, executed respectively by the Filter,

Selector and Input Interface roles (Figure 7).

3.4 Role Interactions Modeling

Each one of the Role Interactions Model shows the

interactions between the internal and external

entities related to each specific goal. The

interactions are numbered according to their

sequencing.

Figure 8 shows the Role Interaction Model related to

the "Model User" specific goal. The Monitor role

captures user navigational behavior. A user profile,

acquired implicitly, is transferred to the “User

Modeler” role so that it can create a user model.

Another alternative is explicit profile acquisition in

which the user explicitly specifies his/her interests

through the “Input Interface” role that sends the

profile to the “User Modeler” role. The alternative

responsibilities of the specific goals determine

alternative interactions between roles. During the

role interactions modeling new variation points are

associated to the specific goals containing variant

groups of role interactions. For the “Model Users”

specific goal (Figure 9) the variation point “Variant

Point Number 2” was created having as variants a

set of role interactions according to the variant

responsibilities associated to this goal.

ICSOFT 2008 - International Conference on Software and Data Technologies

164

Figure 6: Role Model of the “Filter Information” Specific Goal.

Figure 7: Role Model of the “Deliver Recommendation” Specific Goal.

Figure 8: Role Interactions Modeling related to the

“Model Users” specific goal.

3.5 User Interface Prototyping

This phase aims at creating generic screens that will

guide the construction of the user interfaces of the

applications reusing ONTOSERS.

In the login interface (Figure 10) the user

informs his/her login and password in order to be

identified by the respective User Interface agent.

Figure 9: Variability modeling of role interactions

associated to the “Model users” specific goal.

Figure 10: User Interface Prototype (Login).

A CASE STUDY ON DOMAIN ANALYSIS OF SEMANTIC WEB MULTI-AGENT RECOMMENDER SYSTEMS

165

Figure 11: User Interface Prototype (User Identification).

Before logging into the system, the user needs to

specify its profile, which is composed by the

identification and interests of the user. Figure 11

shows the user interface prototype for specifying the

user identification, composed by the name, e-mail,

User interests are organized into categories. The

User Interface Prototype for specifying the interests

within a certain interest category is shown in Figure

12.

Figure 12: User Interface Prototype (Interest Category 1).

The Personalized Recommendation Interface

prototype presents to the user the recommended

items, with their identification and the similarity

with the user model (Figure 13). By clicking on the

information item users can visualize all its content.

Figure 13: Recommendation Interface Prototype.

4 RELATED WORK

A domain model similar to the one described in this

work is presented in Girardi and Marinho (2007).

ONTOWUM-DM is a domain model that describes

the requirements of recommender systems based on

use mining and collaborative filtering. ONTOSERS-

DM extends this model, incorporating the

responsibilities associated to explicit user profile

acquisition, content based filtering and hybrid

filtering. Also, ONTOSERS-DM is based on

Semantic Web information sources. Thus, the reuse

potential of the model is improved by increasing the

variable part of the model thus allowing the

generation of a greater number of recommender

systems based on the different available techniques.

Also, ONTOSERS-DM was developed using a new

version of the MADEM methodology and the

ONTOMADEM tool (Girardi and Leite, 2007,

Girardi and Marinho, 2007).

Semantic Web technologies make use of

ontologies which are a formal and explicit

specification of a conceptualization (Gruber, 1995).

Some works already explore the potential of the

ontologies and Semantic Web technologies for the

improvement of the effectiveness of recommender

systems. Middleton et. Alii (2002) developed an

approach for user modeling based on ontology

inferences, representing the user profiles in terms of

an ontology of scientific articles. In Ziegler (2004)

and Ziegler et al. (2004) is introduced an approach

for hybrid information filtering, that computes the

user profiles and generates recommendations based

on a product taxonomy. A similarity measure

between concepts expressed using description logics

is introduced in Fanizzi et al. (2005). As description

logic is the formalism for ontology representation,

such similarity measure can be applied for

information access in the Semantic Web.

ONTOSERS-DM is a specification of the

requirements of applications based on these

techniques.

5 CONCLUSIONS

This work introduced ONTOSERS-DM, a domain

model that specifies the common and variable

requirements of recommender systems based on the

Semantic Web technology. The variability of the

concepts related to user modeling and information

filtering was described. ONTOSERS-DM was

constructed following the guidelines of the

MADEM methodology with the support of the

ONTOMADEM tool (Girardi and Leite, 2007). The

modeling process revealed being consistent and

capable of generating products with high potential

of reuse.

Currently we have finished the architectural,

detailed design and implementation models of

ONTOSERS, where a solution for the requirements

ICSOFT 2008 - International Conference on Software and Data Technologies

166

specified in ONTOSERS-DM was mapped to a

multi-agent framework.

The ONTOSERS system family was reused in

the construction of a recommender system in the

area of Tax Law, according to the guidelines of the

MAAEM methodology (Lindoso, 2006a, Lindoso,

2006b, Girardi and Lindoso, 2005). Such system is

called INFOTRIB and supports collaborative and

content based filtering. We are currently working on

its extension in order to support also hybrid filtering

techniques.

REFERENCES

Lindoso, A. N., “The SRAMO Technique for Analysis

and Reuse of Requirements in Multi-agent

Application Engineering”. In: IX Workshop on

Requirements Engineering (WER 2006), Proceedings

of IX Workshop on Requirements Engineering, Rio de

Janeiro, 2006a. v. 20. p. 41-50.

Lindoso, A. N., Uma Metodologia Baseada em Ontologia

para a Engenharia de Aplicações Multiagente. Tese

de Mestrado. São Luís. 2006b.

Fanizzi, C., Nicola, D., Espósito, F., “A Semantic

Similarity Measure for Expressive Description

Logics”. Universita di Bari. Italy. 2005.

Ziegler, C., “Semantic Web Recommender Systems”.

Proc. Joint ICDE/EDBT Ph.D. Workshop, pp. 78-89.

2004.

Ziegler, C., Schimitdt-Thieme, L., Lausen, G., “Exploiting

Semantic Product Descriptions for Recommender

Systems”. SWIR 04. ACM SIGIR Semantic Web and

Information Retrieval Woksho, Sheffield, UK, July,

29, 2004.

Adomavicius, G., Tuzhilin, A., “Toward the Next

Generation of Recommender Systems: A Survey of

the State-of-Art and Possible Extensions”, IEEE

Transactions on Knowledge and Data Engineering,

2005 v. 17, n 6. 734-749.

Girardi, R., Leite, A., “ONTOMADEM: An Ontology-

driven Tool for Multi-Agent Domain Engineering”,

Proceedings of The Nineteenth International

Conference on Software Engineering and Knowledge

Engineering, Ed. Knowledge Systems Institute.

Boston, 2007, pp. 559-564.

Girardi, R., Leite, A., “A Knowledge-based Tool for

Multi-Agent Domain Engineering”, Knowledge-Based

Systems Journal, Springer, 2008. (To appear)

Girardi, R., Lindoso, A., “DDEMAS: A Domain Design

Technique for Multi-agent Domain Engineering”. In:

The Seventh International Bi-Conference Workshop

on Agent-oriented Information Systems (AOIS-2005)

at The 24th International Conference on Conceptual

Modeling (ER 2005), Lecture Notes in Computer

Science 3770, Springer-Verlag, Berlin Heidelberg,

2005. p. 141-150.

R. Girardi, L. Drumond, “A Multi-agent Legal

Recommender System” of Ai and Law, 2008 (to

appear).

Girardi, R., Marinho, L., “A Domain Model of Web

Recommender Systems based on Usage Mining and

Collaborative Filtering, Requirements Engineering

Journal”, Ed. Springer-Verlag, London, vol.12, n. 1,

2007, pp. 23-40.

Middleton, S. E., Roure, H. A., David, C., “Exploiting

synergy between ontologies and recommender

systems”. In Proceedings of the WWW International

Workshop on the Semantic Web, volume 55 of CEUR

Workshop Proceedings, Maui, HW, USA, May 2002.

Gruber, T. R., “Toward Principles for the Design of

Ontologies used for Knowledge Sharing”.

International Journal of Human-Computer Studies,

1995. Nº 43, pp. 907-928.

A CASE STUDY ON DOMAIN ANALYSIS OF SEMANTIC WEB MULTI-AGENT RECOMMENDER SYSTEMS

167