ANALYSIS AND DESIGN OF A RECOMMENDER SYSTEM

WITH AGENTS

Javier Portillo-Rodríguez, Aurora Vizcaíno

Alarcos Research Group – Institute of Information Technologies & Systems, Escuela Superior de Informática

University of Castilla – La Mancha, Ciudad Real, Spain

Juan Pablo Soto, Mario Piattini

Alarcos Research Group – Institute of Information Technologies & Systems, Escuela Superior de Informática

University of Castilla – La Mancha, Ciudad Real, Spain

Keywords: Recommender system, Multi-agent system, INGENIAS.

Abstract: The aim of this paper is to provide a guideline for novel multi-agent systems developers in order to assist

them in the development of this kind of systems. Papers frequently focus on describing how systems work,

but seldom describe the different steps carried out to attain the final product. We shall attempt to show how

the usage of a methodology facilitates the analysis, design and implementation phases, along with how the

INGENIAS methodology has helped us to systematically construct a recommender system.

1 INTRODUCTION

Papers related to software agents and multi-agent

systems are often focused on explaining how the

agents work or the goal of the system for which

either the agents or the multi-agent systems have

been designed. However, explanations are rarely

provided of how the methodological development of

this kind of systems should be tackled, i.e. how a

methodology should be used to tackle the analysis

and design phases of a multi-agent system. The

explanation of these methodological aspects might

be of use to future developers who have to confront

the problem of developing these kinds of systems by

using confirmed techniques and methods.

Numerous methodologies focusing on the

development of multi-agent systems currently exist.

This paper will show the steps involved in

developing a multi-agent recommender system. The

first step was to study which was the most suitable

methodology for our problem, finally INGENIAS

methodology was chosen. There now follows a brief

description of the system developed. This is a multi-

agent system whose goal is to recommend

documents in a Community. The system will

therefore offer document recommendation, and will

take into account both the evaluations that the

documents obtain and the trust related to the

evaluators. The reason for using the trust parameter

to make recommendations is that some communities

are often “virtual” as their members may be

geographically distributed. This implies a lack of

face-to-face communication which affects certain

aspects of interpersonal relationships, such as the

possible decrease of trust between members and we

consider that it is highly important to be able to

discover how trustworthy a knowledge source (in

this case are the own members) is. This paper is

structured as follows. The following section explains

the reasons for using the INGENIAS methodology

to develop the recommender system (multi-agent

system). The analysis and design of this system will

be explained in Section 3. Section 4 describes some

implementation aspects in relation to the design

described in Section 3. Section 4 also shows the tool

developed and the recommendation results generated

by this tool. Finally, Section 5 presents our

conclusions and future work.

2 INGENIAS

In INGENIAS, the general approach used to specify

175

Portillo-Rodríguez J., Vizcaino A., Soto J. and Piattini M. (2009).

ANALYSIS AND DESIGN OF A RECOMMENDER SYSTEM WITH AGENTS.

In Proceedings of the 4th International Conference on Software and Data Technologies, pages 175-178

DOI: 10.5220/0002253801750178

 SciTePress

Multi-Agent Systems (MAS) is to divide the

problem into more concrete aspects that form

different views of the system. The difference

between this proposal and others which already exist

is how these views are defined and built, and how

they are integrated into the MAS development. Each

type of view is described by using a meta-modelling

language which specifies a kind of grammar with

which models are built, known as views. INGENIAS

recognizes five meta-models that describe system

views.

After studying other methodologies, we

developed an analysis with which to identify the

advantages of INGENIAS in comparison with the

other methodologies. First, we concluded that

INGENIAS provides the best support in all the

phases of the lifecycle of an agent-based application,

including its management. Second, INGENIAS

provides a tool support that others do not yet have.

For instance, INGENIAS provides a visual language

for multi-agent systems definition, which is a

process to guide the lifecycle of software

development based on the Rational Unified Software

Development Process (RUP). In addition, the code

generator itself is another tool developed for

INGENIAS. It produces codes according to the

instructions given by the specification of the system.

This tool can be used to generate a code no matter

what agent platform is being used (Pavón and

Gómez-Sanz 2006).

The analysis and design of the recommender

system mentioned in the introduction is described in

the following sections.

3 ANALYSIS AND DESIGN WITH

INGENIAS METHODOLOGY

Before describing the analysis and design it is first

necessary to give a detailed description of the

system that it is going to be developed, in this case a

multi-agent system that must be able to recommend

documents in a comunity.

Software agents must recommend documents to

users that will prevent them from making costly

searches which, on many occasions, do not generate

the expected results. In these communities it is

important to know which users are trustworthy

document sources. A trust model presented in

(Portillo, Soto et al. 2008) is therefore used to make

recommendations and to calculate the trust between

members. The model then uses these trust values to

weight the evaluation values given by users to the

documents once they have used them. The

evaluations made by trustworthy users will therefore

have more importance than the evaluations made by

non-trustworthy users.

Once the kind of system to be developed is

known, system requirements must be extracted.

Moreover, the design requirements specify that each

user is represented in the system by one software

agent and that the communities will be managed by

another software agent.

These requirements are the starting point of the

INGENIAS analysis phase in which requirements

are used to identify use cases that will be modelled

by using interaction models such as the use case

Obtain Recommendation.

Figure 1: Obtain_Recommendation Use case.

Figure 1 shows an example of the modelling of

the Obtain_Recomendation use case through the

use of an interaction model (called

document_recommendation). In this model the goal

of the use case is to Recommend Documents. The

roles that participate in the use case are also

specified, indicating which one initiates the use case

and which one collaborates. In the case of Figure 1

there are three roles: the user (played by a human

being), the recommender and the community

manager (both of which are played by software

agents). It is important to emphasize that these

kinds of diagrams are previously modelled by using

UML collaboration diagrams, and this is indicated

by using the HasSpec Relationship.

The task of specifying use cases by using

interaction models must be carried out for each of

the use cases previously identified. In this case, only

one use case model (that which is most closely

related to the main goal of the system

(recommending documents) is mentioned owing to

space limitations.

ICSOFT 2009 - 4th International Conference on Software and Data Technologies

176

Figure 2: Agent Model for User Agent.

To specify which agents will interact in the

systems and which tasks they will carry out it must

be used what is termed as the ‘agent model’ in

INGENIAS. These models must be used at the

beginning of the analysis phase to identify which

kind of agents must be considered in the multi-agent

system but, in this case, the agents were identified

with the design requirements and these models have

therefore not yet been used.

As was mentioned in the design requirements,

one type of software agent is needed to represent

users in the communities which is able to

recommend documents (User Agent) (see Figure 2),

and another agent is needed to manage community

information (Manager Agent)(not included due to

space limitations).

Figure 2 therefore shows that the User Agent has

the goal of Recommend Documents and it can play

the roles of community member or recommender.

This analysis phase is completed by representing

all the identified elements, including the agents, in

an organization model which will represent the

architecture of the multi-agent model.

One of the most important elements that must be

incorporated into the architecture through the

organization model are WorkFlows (WFs). WFs are

sets of tasks created by grouping the tasks

previously identified during the analysis phase.

Figure 3: Agent Model for Manager Agent.

WFs are the starting point of the design phase.

Two main WFs have been identified in this system:

 Community Management WF: This

includes tasks related to community

management such as registering, connections,

etc.

 Documents Management WF: This WF is

specialized in three SWFs: Add Documents,

Documents Evaluation, Documments

Recommendations (see Figure 3).

Each WF can be divided into different Sub-

WorkFlows (SWFs) by analysing which tasks might

be necessary to complete the SWF. These tasks will

be carried out by the agents.

Once the tasks that are related to each SWF have

been identified it is important to establish the

sequence of execution, that is, it is vital to indicate

which task will be the first to be executed, which

will be the second, etc.

Moreover, when setting this sequence, it is

necessary to include all the elements (in INGENIAS

these are called Facts) that are needed for each task

to be executed and those elements

produced/generated for each task after being

executed.

It is necessary to specify not only the sequence

of execution but also the interactions generated

between agents when executing these tasks in each

SWF. This activity can be modelled by using the

INGENIAS Interaction Models. These models

specify each interaction by identifying a set of

Interaction Units that represent each occasion on

which an agent communicates with another agent.

By completing this specification for each WF

sufficient detail can be obtained to translate these

models to the implementation of the multi-agent

system by using any agent platform.

The tool constructed by using the previously

explained multi-agent system is also shown.

4 IMPLEMENTING THE

RECOMMENDER SYSTEM

The main idea behind translating the design of the

multi-agent system to implementation in JADE is to

consider each SWF as a JADE Behaviour. That is, in

each SWF there is a set of tasks, and some of these

tasks are executed by one agent and others are

executed by another agent. The intention is to codify

each set of tasks that an agent executes in each SWF

into a Behaviour. Thus, each agent will carry out

each SFW by executing the corresponding

ANALYSIS AND DESIGN OF A RECOMMENDER SYSTEM WITH AGENTS

177

Behaviour when necessary. Using this idea we have

developed our tool. This tool allows users to search

communities in order to register or connect and ask

for recommendations. Moreover, the user can obtain

information regarding all the active communities to

discover which might be of interest to the user.

Therefore, once the user is connected to a

community there are several possibilities. The user

can send documents to the community by accessing

the Contribution Menu or solicit documents related

to a topic by accessing the Recommendations Menu.

Figure 4 shows the results of a recommendation

where in the left part of the screen there is a list of

documents recommended and on the right there is

another list with the documents that should be

evaluated.

Figure 4: Recommendation results.

In order to evaluate the tool we are carrying out a

performance test. The test is oriented towards the

Manager Agent’s performance. There is only one

Manager Agent per Community and it is important

to discover whether this agent is able to manage all

the incoming messages from User Agents when a

high number of users are connected to a Community.

The preliminary results indicate that the tool starts to

become less efficient when there are more than 800

users consulting at the same time. However, as we

recommend this tool to small/medium-sized

companies it is expected that the number of users

will be less than 800 and that the tool will work

efficiently.

5 CONCLUSIONS AND FUTURE

WORK

This paper shows how to confront a multi-agent

system development by using a systematic method

which, in this case, was the INGENIAS

methodology. In our case INGENIAS was chosen

since it is a complete methodology which is well

supported by a design tool. The only problem is that

INGENIAS involves the generation of a

considerable amount of documentation which is

sometimes redundant.

We have explained how to carry out the activities

proposed in INGENIAS in the analysis and design

phases with the goal of assisting future multi-agent

system developers to discover which steps to follow

when developing a system. We have also described

the different models used for developing a

recommender system.

As future work, we intend to explore the

possibility of automatic code generation that

INGENIAS offers through the latest versions of its

Ingenias Development Kit (IDK).

ACKNOWLEDGEMENTS

This work is partially supported by the MELISA

project (PAC08-0142-3315) y ENGLOBAS project

(PII2I09-0147-8235) Junta de Comunidades de

Castilla-La Mancha, Consejería de Educación y

Ciencia, in Spain and partially supported by the

ESFINGE project (TIN2006-15175-C05-05)

Ministerio de Educación y Ciencia (Dirección

General de Investigación)/Fondos Europeos de

Desarrollo Regional (FEDER) in Spain, and

CONACYT (Mexico) under the third author’s

scholarship grant 206147.

REFERENCES

Pavón, J. and J. Gómez-Sanz (2006). INGENIAS web site.

http://grasia.fdi.ucm.es/ingenias/.

Portillo, J., J. P. Soto, et al. (2008). A Model to Rate Trust

in Communities of Practice. International Conference

on Enterprise Information Systems (ICEIS 2008),

Barcelona, Spain.

ICSOFT 2009 - 4th International Conference on Software and Data Technologies

178