AUTOMATED PRODUCT RECOMMENDATION BY EMPLOYING

CASE-BASED REASONING AGENTS

Ozgur Baykal

Reda Alhajj

Faruk Polat

Dept of Computer Engineering, Middle East Technical University, Ankara, Turkey,

ASDA Lab, Dept of Computer Science, University of Calgary, Calgary, AB, Canada,

Keywords:

Multiagent Systems, Case-Based Reasoning, E-Commerce, Product Recommendation.

Abstract:

This paper proposes a cooperation framework for multiple role-based case-based reasoning (CBR) agents to

handle the product recommendation problem for e-commerce applications. Each agent has different case struc-

ture with intersecting features and agents exploit all information related to the problem by cooperation, which

is accomplished through the merge of distributed cases. The role-based CBR agents merge the distributed

cases by introducing a global heuristic function, which exploits the relevancy of each merged case within the

viewpoint of each agent and the satisﬁed/unsatisﬁed problem constraints. Finally, the proposed framework has

been tested for elective course recommendation.

1 INTRODUCTION

In this paper, we propose a cooperative method that

improves solution quality of CBR agents by the coop-

erative retrieval of distributed cases. An agent consti-

tutes an autonomous node of the respective distributed

case-base. The cooperation of agents is accomplished

by merging distributed cases to form overall cases

having better representation of the problem. This way,

a problem is cooperatively solved by a society of au-

tonomous agents in order to improve solution qual-

ity (Sycara 1998). Further, the cooperation of agents

results in having each agent uses its expertise to re-

trieve distributed cases and to evaluate the overall

cases that will be used in the solution. They negoti-

ate on possible results of merging distributed cases to

locate erroneous ones according to a heuristic func-

tion, which is used for directing the result of merge

in increasing the relevancy of overall cases and for

handling noisy distributed case-bases. So, the most

salient property of the work described in this paper is

the merge of distributed cases by autonomous agents

guided by heuristic functions in order to introduce

efﬁcient evaluation of the resulting overall cases by

considering the limited viewpoint of each agent, and

to provide a robust way of merging distributed case-

bases with noisy cases, by negotiation between agents

over the conﬂicting feature values. The rest of the pa-

per is organized as follows. The proposed multiagent

CBR framework is described in Section 2. Experi-

mental results for elective course recommendation are

discussed in Section 3. Section 4 includes summary

and conclusions.

2 PROPOSED COOPERATION

FRAMEWORK

The overall case structure consists of the following

values: 1) feature-agent(s) tuple, denoting agents re-

lated to a problem feature; it holds the values of prob-

lem features; 2) set of problem constraints, includes

constraints of the current problem satisﬁed and un-

satisﬁed by the overall case; 3) value of information

of the overall case within the limited viewpoint of

each agent; 4) value of coherency of the overall case,

denoting the consistency of a component within the

view point of the relevant agent. Finally, the relevance

of the overall cases can be determined from the value

of information, value of consistency, and the satisﬁed

constraints.

The overall cases are stored in an overall case-base

(OCB) shared by all agents. Every time a problem is

presented, a new OCB is generated to be used by the

agent working on the problem for a global case-base

in the context of the current problem. Since the struc-

ture and platform of the running agents are mostly dif-

ferent, OCB has been implemented as an agent with

communication capabilities. It is controlled by an

agent called OCB-interface.

Consider an agent a

and let e

be its evaluation

515

Baykal O., Alhajj R. and Polat F. (2004).

AUTOMATED PRODUCT RECOMMENDATION BY EMPLOYING CASE-BASED REASONING AGENTS.

In Proceedings of the Sixth International Conference on Enterprise Information Systems, pages 515-518

DOI: 10.5220/0002620105150518

 SciTePress

function; α

be the set of all feature-value pairs of

overall case α that can be evaluated within the view-

point of a

, i.e., the values of features in the agent’s

local case structure; and g

be a distance function

deﬁned over distributed cases of agent a

. An eval-

uation function should be fast and simple because it

will be used after each merge on the overall case.

Distance functions are percentage functions deﬁned

via the ratio between the values of features of a dis-

tributed case; they are simple and fast.

Satisﬁed constraints have the value 1, and unsatis-

ﬁed constraints have the value 0. The semantics of

this heuristic function are used in constructing over-

all cases. The overall value shows the relevancy of

an overall case to the problem. The overall value of a

case is the evaluation of the following points within

the viewpoint of the system: 1) the solution qual-

ity within the viewpoint of each agent (value of in-

formation); 2) the percentage of the change of the

originating distributed case by the sequence of merge

operations (value of consistency); 3) the set of sat-

isﬁed/unsatisﬁed constraints. The overall evaluation

function of the agent working on the problem is trans-

mitted to other agents as expressions. This way,

agents share the overall evaluation function in order

to evaluate overall cases.

Each agent uses its limited viewpoint to retrieve

from its case-base the most promising local cases to

be used as initial seeds in initiating the merge of dis-

tributed cases. When all the distributed cases are not

used as initial seeds, the combination of initial seeds

with the distributed cases may not lead to the best

overall cases. However, initiating all cases will signif-

icantly decrease the response time of the application

when the number of cases is very high.

Retrieved initial seeds are inserted into the over-

all case base. The value of the information features

of the overall case are requested from the relevant

agents. The request is only possible when more than

one agent share the same case-structure. Since each

agent is capable of solving the problem individu-

ally, they manipulate the constraints available in their

viewpoints for checking. So, the satisﬁed/unsatisﬁed

constraints of the initial seeds are set by the agent re-

trieving the initial seeds.

Each agent selects some local cases per overall case

using its selection function. Selection is a uniﬁcation

of the set of intersecting features of an overall case

and the agent’s local case as well as differing fea-

tures of agent’s problem and the overall case. When

distributed cases are noisy or have missing parts, the

merge operation results in inconsistent cases.

Each possible result of merge, called temporary

overall case, is stored in the agent’s temporary over-

all case table, denoted temp-OC. Constraints of the

temporary overall cases are checked by the agent per-

forming the merge. Then each relevant agent is in-

formed about the changes of the temporary overall

case for the evaluation of the value of information and

value of consistency. Selecting the temporary overall

case with the highest evaluation value for the result

of the merge makes the search climbs through hills

to some local minima. The selected merge result is

only inserted in the overall case-base when the overall

evaluation value of the selected overall case is higher

than the overall evaluation value of the unmerged

overall case (merge if promising principle). There-

fore, merge of overall cases and distributed cases may

not lead to complete cases.

The revision of cooperative retrieval can be accom-

plished by the revision done in the retrieval of initial

seeds and the selection of distributed cases for merge.

One of the possible ways of accomplishing the revi-

sion is restarting the cooperative retrieval by remov-

ing the distributed components of the overall cases

which are used by selected distributed cases for merge

and by retrieved distributed cases as initial seeds.

The overall case-base is implemented using OCB-

interface module, which is instantiated by the agent

that presents the problem. An overall case-base is

generated per active problem using the OCB-interface

by the agent to whom the problem has been submit-

ted. It has communication capabilities with the agents

and has own control over the overall cases in order to

provide safe computations on them.

An agent holds the problem it is working on, a list

of registered agents, a message queue for storing mes-

sages, retrieved cases used for initiating merge opera-

tions, a temp-OC table for holding temporary results

of merge operations it is currently working on, and

selected cases for the possible merges of an overall

case with distributed case-base. Since more than one

problem is to be solved by the system, the problems

are distinguished with a unique label referring to the

agent to whom the problem is presented. The list of

registered agents enables the communication with the

other agents, and each agent needs to register itself

in order to communicate using messages. The mes-

sage queue provides asynchronous computations of

the agents, such that they can continue their execution

after placing their messages into the queue.

Except announcing the problem and the overall

evaluation function, most of the time agents commu-

nicate with each other in order to request information

or as a response to a request. The information trans-

mitted between agents is the value of information and

value of coherency of the overall cases. If a change

on the features that are related with agents occurs af-

ter the merge operation, agents request the value of

information and value of coherency through transmit-

ting relevant distributed components. Agents check

constraints using their available information, since all

distributed components of an overall case are checked

for satisfaction by the relevant agents. These pro-

ICEIS 2004 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS

516

cesses need to be synchronized because overall cases

are merged with selected distributed cases and dis-

tributed cases are selected per overall case.

The information passing mechanism in the sys-

tem is peer-to-peer, although the OCB-interface is

considered as a shared memory for holding overall

cases. Since all overall cases are maintained by OCB-

interface, it becomes bottleneck in the communica-

tion infrastructure. Despite that OCB-interface con-

trols overall cases without any overhead, it proposes

no solution in the case of failure.

The messages used in the system for communica-

tion are KQML messages (Finin, McKay and Fritson

1992). This way, all the information related with the

content of the message and with ontology of the prob-

lem could be sent along with appropriate message for-

mat. KQML provides separation of semantics of the

message from the communication protocol. There-

fore, the use of KQML in the system enables different

structured agents to cooperate.

3 EXPERIMENTS

The elective course recommendation problem has

been selected for evaluating the proposed framework.

The environment on which agents operate is a subset

of the Student Affairs database of Middle East Tech-

nical University. Experimental results and the elective

recommendation problem with hypothetically gener-

ated cases are presented in relation with the product

recommendation problem.

The elective course recommendation problem is a

classiﬁcation problem. The superset of the set of of-

fered courses and the set of all students constitute the

domain of the problem, and the set of courses that can

be taken as electives constitutes the range of the prob-

lem. There are also inter-dependencies between prob-

lem features that contribute to the potential solution;

these are cumulative grade point average (CGPA), list

of previously taken elective courses, and

OSS

scores

of the student. The experimental system consists

of three agents, namely CGPA-Agent, ELIST-Agent

and

OSS-Agent. Each agent controls the cases de-

termined by its role at the respective distributed node.

The distributed cases of each agent are constructed by

the concrete enrollments of a student. Note that dis-

tributed cases may be redundant, and redundant cases

are subsumed by the more powerful cases (Yang and

Racine 1998).

A student can take an elective course according to

the preferences of CGPA, previously selected courses,

and

OSS scores using CGPA-Agent, ELIST-Agent

OSS is a university entrance exam that measures verbal

and analytic skills of each high school student in Turkey.

and

OSS-Agent, respectively. One of the preferences

determines the agent to solve the problem, and the

selected agent completes the following processes to

solve the problem using CBR: 1) cooperative retrieval

of its cases in order to exploit all problem relevant in-

formation; 2) adaptation of selected case(s) of coop-

eratively retrieved cases to generate an answer; 3) re-

vision of cooperative retrieval in case of failure.

We tested the number of messages and the per-

centage of complete cases with respect to the change

in distributed case-base size and number of retrieved

cases for 3-agent scenario. Each time, the obtained

experiment results are averaged over randomly se-

lected 100 problems, and one distributed case is se-

lected with each overall case by each agent. The fol-

lowing heuristic function h is used in the experiments,

where n is the number of agents in the system.

h(V I

, V C

, ..., V I

, V C

) =

i=0

(V I

∗ V C

)

(a) Number of messages (b) Percentage of complete cases

Figure 1: 3-Agent Scenario with increasing number of re-

trieved initial seeds

The results for increasing number of retrieved ini-

tial seeds and 200 distributed case-base size are given

in Figure 1. Figure 1-a shows the average number

AUTOMATED PRODUCT RECOMMENDATION BY EMPLOYING CASE-BASED REASONING AGENTS

517

(a) Number of messages (b) Percentage of complete cases

Figure 2: 3-Agent Scenario with increasing size of dis-

tributed case-bases

of messages transmitted between the agents with re-

spect to the number of initial seeds. From Figure 1-

a, it can be easily seen that the ratio of the number

of messages transmitted to the number of retrieved

initial seeds is constant. Figure 1-b shows the per-

centage of complete cases with respect to the num-

ber of initial seeds. The percentage of complete cases

decreases in Figure 1-b because the more distributed

cases are used, it is more probable to have distributed

cases with higher relevance, so that the possible re-

sult of the merge operation mostly has lower heuristic

value. In other words, the result of the merge opera-

tion has lower heuristic value most of the time.

The results for increasing size of distributed case-

bases and 10 retrieved initial seeds are given in Fig-

ure 2. Figure 2-a shows the average number of mes-

sages transmitted between the agents with respect to

the size of distributed case-bases. It can be easily seen

that the number of messages transmitted tend to de-

crease with increasing number of messages. Figure 2-

b shows the percentage of complete cases with re-

spect to the size of distributed case-bases. According

to Figure 2-b, the percentage of complete cases de-

creases slightly with the increasing size of distributed

cases. Note that agents can exploit subset of the prob-

lem relevant information for partial cases because

agents can solve problems for non-complete cases.

It has been observed from the experiments that

when the size of distributed cases is huge, the num-

ber of messages is quite admissible with respect to

the size of distributed case-base. Therefore, the pro-

posed framework can be efﬁciently used with physi-

cally distributed case-bases. Finally, the percentage of

complete cases is low because they are not used in the

generation of distributed cases, and the distribution

of the evaluation function (for evaluating the value of

information) and distance function (for evaluating the

value of consistency) of each agent are different.

4 SUMMARY AND

CONCLUSIONS

The proposed approach is simple and provides com-

putationally efﬁcient cooperation framework for re-

trieval of distributed case-bases. It is applicable to

systems having small number of agents in which the

distributed case-bases are physically distributed and

in which the number of retrieved initial seeds and the

number of selected distributed cases for merge are

small. Since role-based CBR agents have limited de-

scription of the problem, they need to cooperate in

order to improve the solution quality. The retrieval

from distributed case-bases results in sub-optimal re-

trievals because of the fact that the construction of

overall cases highly depends on the initiated cases and

the progressing heuristic values of the merged overall

cases. The use of heuristics in the merge of distributed

cases makes the search as an application of hill climb-

ing search in broadening hyper-planes. The merge of

distributed cases handles noisy distributed cases by

negotiation of agents on the possible results of merge.

Finally, the proposed cooperation framework can be

extended with cooperative application of adaptation

functions within the multiagent framework.

REFERENCES

Finin T., McKay D. and Fritson R., “An overview of

KQML: A Knowledge Query and Manipulation Lan-

guage,” Technical Report, University of Maryland,

1992.

Yang Q. and Racine K., “Redundancy and Consistency De-

tection in Large and Semi-Structured Case-Bases,”

Technical Report, 1998.

Sycara K., “Multiagent Systems,” AI Magazine, Vol.19,

Vol.2, 1998.

Wooldridge M., “Intelligent Agents,” in Multiagent Sys-

tems, Weiss G. (Ed.), MIT Press, 1999.

ICEIS 2004 - ARTIFICIAL INTELLIGENCE AND DECISION SUPPORT SYSTEMS

518