MULTI-AGENT SYSTEMS IN DATA IMPUTATION

OF COLLABORATIVE FILTERING

In Case of e-WeddingThailand

Kunyanuth Kularbphettong

Science and Technology Faculty, Suan Sunandha Rajabhat University, Dusit Bangkok, Thailand

Phayung Meesad, Gareth Clayton

Information Technology Faculty, King Mongkut’s University of Technology North Bangkok, Bangkok, Thailand

Keywords: Collaborative filtering, Multi-agent system, e-WeddingThailand, Web services, Ontology, Data imputation,

Data preprocessing techniques.

Abstract: Multi-agent system is rapidly emerging as a new paradigm to develop complex and intelligent commerce

application systems in e-Business. In this paper, we present the findings on the techniques used for data

imputation techniques in Collaborative filtering based on multi-agent systems (MAS) of the on-going

project, e-WeddingThailand. The aim of our project is to implement MAS combined with various

techniques, like Web Services, Ontology, Web Semantic and Data Mining techniques. However, the present

paper focuses on the data imputation technique in collaborative filtering utilized in order to treat missing

values of customer behavioral patterns for Wedding business. As a result, a model obtained is therefore used

as a benchmark for testing potential patterns so that they are used to strengthen the derived model in

enhancing the overall system performance.

1 INTRODUCTION

Collaborative Filtering is a well known technique

used in recommendation systems to suggest products

and services for customers on e-Commerce systems

because it is highly rated by some other customers

with similar tastes. This technique generally

analyzes relationships between users and products or

services to identify the user product/service

associations (Yifan, et al., 2008). However, the

collaborative filtering technique has some

disadvantages, like the inaccuracy of prediction and

the lack of the transparency in the predictions

(Janusz, 2006), when the number of similar users is

small and the collected data from customers contain

many missing values.

Generally, missing values can generate a bias

impacting on the quality of the performance of data

mining/machine learning techniques. The more

incorrect and irrelevant data presented, the more

unreliable the results. Especially in collaborative

filtering technique, missing data impacts on the

difficulties in estimation and inference (Schafer,

1997). Therefore, Missing Data Imputation is one of

the challenging issues that impacts on improving the

performance of Data Mining and Machine Learning

approaches.

Also, with the growth and success of the Internet,

there are tremendous numbers of customers using

recommendation systems to rapidly perform

adequate and reliable results in order to meet the

demands of customers. With widely developed

advanced mechanisms, Multi-agent System is one of

the promising approaches utilized to enhance the

performance of the recommendation systems. A

multi agent system is a computational system, or a

loosely coupled network in which two or more

agents interact or work together to perform a set of

tasks or to satisfy a set of goals. Each agent is

considered as a locus of a problem-solving activity

which operates asynchronously with respect to the

other agents (Sandholm and Lesser, 1996). Although

there has been an enormous amount of research done

recently applied MAS in a wide range of problem in

Data Mining and Machine Learning techniques,

457

Kularbphettong K., Meesad P. and Clayton G..

MULTI-AGENT SYSTEMS IN DATA IMPUTATION OF COLLABORATIVE FILTERING - In Case of e-WeddingThailand .

DOI: 10.5220/0003186304570461

In Proceedings of the 3rd International Conference on Agents and Artiﬁcial Intelligence (ICAART-2011), pages 457-461

ISBN: 978-989-8425-41-6

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

there is little of research focused on using MAS in

the data preprocessing step. Therefore, in this paper

we propose and implement the MAS framework of

an on-going project, called e-WeddingThailand,

which merely emphasizes using MAS to handle the

problems of missing values in Collaborative filtering

techniques.

The remainder of this paper is organized as

follows. Section 2 reviews related literature and

research works. Section 3 presents the experimental

results based on the purposed framework based on

multi-agent systems. This prototype demonstrates

how to succeed in adapting multi-agent systems in

data imputation in collaborative filtering techniques.

Finally, we conclude the paper with future research

issues in section 4g.

2 RELATED WORKS

A literature search shows that most of the related

research has been deployed multi-agent to develop

e-Commerce in collaborative filtering techniques

and the data imputation techniques by following

this:

According to Wei et al, 2003, they present a

market –based recommendation system based on

MAS where an agent acts on behalf of its users and

sells the recommended slide bar space. Also, Gulden

Uchyigit and Keith Clark, 2003, applied an agent

based approach to collaborative filtering and used a

process of pre-clustering to form shared interest

based on their interest profiles. MANET (Collins et

al, 2002) a mobile agent based system, enables users

to compare products from different sellers.

However, it has a problem of system performance

when the number of sellers increases. Furthermore,

according to Zili Zhang and research group

, 2003,

they applied agent technology to collect and

integrate data distributed over various computing

platforms to facilitate statistical data analysis in

replacing the missing values by using either

Approximate Bayesian Bootstrap (ABB) or Ratio

Imputation and using MAS improves the execution

time by focusing on spatial knowledge in order to

extract knowledge in Predictive Modeling Markup

Language (PMML) format (Baazaou et al, 2005).

Moreover, other research works show that agent

technologies representing in various ways that are

related with data mining techniques (Huang et al,

2007), (Zili et al, 2003) and (Chien et al, 2009).

Hence, from previous literature, it presents that there

are many research studies exploiting multi-agent

technology in collaborative filtering and data mining

techniques blended together. Consequently, in order

to succeed on their frameworks, agents should have

abilities to perform as a behalf of user to handle with

given tasks such as planning, reasoning and learning.

Also, data mining techniques are an important way

to make a reason for agent under uncertainty and in

incomplete information situations. Notwithstanding,

data imputation technique acts as a crucial task in

enhancing performance of the system when mining

incomplete data.

3 THE PROPOSED

FRAMEWORK AND

EXPERIMENTAL RESULTS

This section displayed the experimental results in

data imputation of collaborative filtering of the

e-WeddingThailand project and also compares the

results between the original data set and the imputed

data set.

3.1 The Proposed Framework

For illustration of framework as figure 1 (Kobkul,

2007), and (Kobkul et al, 2010), we select the

wedding businesses and its environment. Nowadays,

the wedding business is increasingly becoming

growth. Booming wedding businesses affect other

businesses like Hotels, wedding studios, car hiring

companies, flower shops, music shops, travel

agencies, and even media businesses. Due to the

explosion of internet, couples use the wedding portal

sites as a medium to search for needed information.

However, it takes timing because of the

overwhelming information available on the internet.

Couples must take time to looking for wedding

packages and related services, and rely on

information of wedding business agencies. Further,

couples make their decision based on comparing

various wedding packages. The wedding package is

composed of hotel wedding package, wedding studio

package, music, floral decoration, card & gift

agencies and etc.

Also, the proposed architecture of the e-

WeddingThailand, this framework is concerned

about four main aspects: multi-agent system, Web

services, ontology and data mining techniques. In

the multi-agent system, each agent is autonomous to

be able to make decisions and act proactively.

Agents can communicate, exchange knowledge,

collaborate or negotiate with each other, to

efficiently achieve the common goal. They receive

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

458

and process users' requirements, make use of user’s

preference, perform as a mediator of services

business. They select and contact with appropriate

wedding business agencies like hotels, wedding

studio, and so on through Web services and give

couples the optimal result. Web services define,

provide related services, and interact with negotiator

agents and ontology is the meaningful data which

can be directly accessed by agents or people through

the Web.

Figure 1: The purposed architecture of the

e-WeddingThailand system.

Figure 2: The rating score web page in e-

WeddingThailand System.

Figure 2 shows the rating score web page of e-

WeddingThailand and figure 3 presents the

operation of JADE in this project. However, in this

paper we present merely a multi-agent system

dealing with data imputation of collaborative

filtering techniques.

Figure 3: The operation of JADE in e-WeddingThailand

System.

3.2 The Experimental Result

We worked on the expected customer data set from

this project where users entry and rated the relevant

wedding products and services and we were able to

collect data on about 315 users. All data was

collected in accordance with appropriate end user

agreements and privacy policies. The analysis was

done with fully anonymous data. It means no

personally identifiable

information collected in

connection with this research. We collected some

relevant wedding products and services. The training

data contains R

values, for each user u and item i,

which represent how user u rated, item i. In addition,

we use a similarly constructed test set. Our system is

trained in order to generate predictions of what users

will prefer the related wedding products and services

and we evaluate the quality of the different

algorithms for computing the predictions using the

mean absolute error (MAE). It is commonly used to

measure the performance of prediction.

MAE









(1)

where p

is the user-specified score

and q

is the calculated score from the system

In our experiment, we filled missing data in the

rating table by implementing two different

algorithms KNN Imputation and the composite

imputation method between hot deck and nearest

neighbour methods based on mean substitution as

described in section 3 and we tested them on our

data sets.

MULTI-AGENT SYSTEMS IN DATA IMPUTATION OF COLLABORATIVE FILTERING - In Case of

e-WeddingThailand

459

Table 1: MAE Performance of CF without imputed and

CF algorithms with KNN Imputations.

MSE of CF

without

Imputed

MAE of CF with KNN Imputation

n=10 n=20 n=30 n=40 n=50

0.768 0.763 0.761 0.758 0.755 0.753

0.776 0.773 0.771 0.769 0.763 0.759

0.753 0.751 0.749 0.743 0.741 0.738

0.759 0.754 0.751 0.748 0.744 0.741

0.761 0.759 0.756 0.753 0.749 0.747

Figure 4: MAE Performance of CF algorithms with KNN

Imputations.

Table 2: MAE Performance of CF without imputed and

CF algorithms with the composite imputation method

between hot deck and nearest neighbour methods.

MAE of CF

without

Imputed

MAE of CF with the composite imputation

method between hot deck and nearest

neighbor methods

n=10 n=20 n=30 n=40 n=50

0.768 0.761 0.758 0.753 0.751 0.747

0.776 0.767 0.763 0.759 0.753 0.749

0.753 0.750 0.747 0.744 0.741 0.738

0.759 0.755 0.751 0.747 0.742 0.739

0.761 0.758 0.755 0.751 0.748 0.745

Figure 5: MAE Performance of CF algorithms with the

composite imputation method between hot deck and

nearest neighbour methods.

After computing for each imputation algorithm,

we performed them with the collaborative filtering

algorithm with the neighbourhood and used weight

to generate the prediction and tested these

experiments with MAE. Table 1 and 2 show the

experimental results. There are five data sets used in

this experimental in which 80 % of data set was of

the training and 20% was of the tested set. And

figure 4 and 5 present the MAE performance of CF

algorithm with KNN Imputations and with the

composite imputation method between hot deck and

nearest neighbour methods.

Nevertheless, from the experiment, we found that

the value of MAE performance of CF algorithms

with the composite imputation method between hot

deck and nearest neighbour methods was lower than

the value of MAE with other approaches. It is clear

that using imputation approaches helps to reduce the

error of the prediction results. However, it takes time

consuming to run the experiment which is shown in

figure 6.

Figure 6: Recommendation Time to user.

4 CONCLUSIONS

In this paper we presented our preliminary ideas of

building multi-agent systems in data imputation of

collaborative filtering based on e-WeddingThailand

system. In the part of MAS, we have implemented

this prototype by using JADE platform. JADE is

quite easy to learn and use. Moreover, it supports

many agent approaches such as agent

communication, protocol, behaviour and ontology.

As for the future work, we need to explore more

reasonable data mining technologies.

ACKNOWLEDGEMENTS

The authors would like to thank Suan Sunandha

Rajabhat University for scholarship support this

project.

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

460

REFERENCES

Chien-Ho Wu, Yuehjen E. Shao, Jeng-Fu Liu, and Tsair-

Yuan Chang, 2009. “On Supporting Cross-Platform

Statistical Data Analysis Using JADE”, Book Series

Studies in Computational Intelligence, Springer

Berlin Heidelberg, issn1860-949X (Print) 1860-9503

(Online), Volume 214/2009.

Gulden Uchyigit and Keith Clark, 2003. “A multi-agent

Architecture for Dynamic Collaborative Filtering”,

Proceedings of the 5th International Conference on

Enterprise Information Systems, Angers, France, April

22-26, pp363-368.

Huang Xin Li; Chosler, R., 2007. “Application of

Multilayered Multi-Agent Data Mining Architecture to

Bank Domain”, Wireless Communications,

Networking and Mobile Computing .International

Conferenceon Volume, Issue , 21-25 Sept. pp 6721 –

6724.

H. Baazaoui Zghal, S. Faiz, and H. Ben Ghezala, 2005. “A

Framework for Data Mining Based Multi-Agent: An

Application to Spatial Data”, World Academy of

Science, Engineering and Technology.

Janusz Sobecki, 2006. “Implementations of Web-Based

Recommender Systems Using Hybrid Methods”,

International Journal of Computer Science &

Applications, 2006 Vol.3 issue 3 pp 52-64.

J. Collins, W. Ketter and M. Gini, 2002. A Multi-agent

Negotiation Tested for Contracting Tasks With

Temporal and Precedence Constraints, Int’l Journal of

Electronic Commerce, pp 35-57.

Kobkul (Kunyanuth) Kularbphettong, 2007. “e-

Negotiation based on Multi-agent ssyetm”,JCSSE

2007 –The international Joint Conference on

Computer Science and software Engineer,Thailand.

Kobkul (Kunyanuth) Kularbphettong, Gareth Clayton, and

Phayung Meesad, 2010. “e-Wedding System based on

Multi-System”, Advances in Intelligent and Soft-

Computing ,series of Springer.

Kobkul (Kunyanuth) Kularbphettong, Gareth Clayton, and

Phayung Meesad, 2010. “A Hybrid System based on

Multi-agent System in Data Preprocessing stage: e-

Wedding”, International Journal of Computer Science

and Information Security.

Sandholm, T. and Lesser, 1996. “Advantages of a Leveled

Commitment Contracting Protocol.” Thirteenth

National Conference on Artificial Intelligence (AAAI-

96), pp. 126-133, Portland, OR.

Schafer, J. L., 1997. Analysis of Incomplete Multivariate

Data, New York: Chapman and Hall.

Wei Y., Moreau L., Jennings N., 2003. Recommender

Systems: A Market-Based Design, Proceedings of the

Second International Joint Conference on Autonomous

Agents and Multi-agent Systems, ACM Press, 600-

607.

Yifan Hu, Yehuda Koren,and Chris Volinsky, 2008.

“Collaborative Filtering for Implicit Feedback

Datasets”, Proceedings of the 2008 Eighth IEEE

International Conference on Data Mining, Pages: 263-

272.

Zili Zhang, Chengqi Zhang and Shichaozhang, 2003. “An

Agent-Based Hybrid Framework for Database

Mining”, Applied Artificial Intelligence, 17:383–398.

MULTI-AGENT SYSTEMS IN DATA IMPUTATION OF COLLABORATIVE FILTERING - In Case of

e-WeddingThailand

461