A Framework of Agent-Based Personalized Recommender System for E-Commerce
Srikumar Krishnamoorthy · South Asian Journal of Management · 2004
The advent of electronic commerce provided a variety of opportunities for customers in selecting products of their choice. But, customers are often overwhelmed by the information overload before they select the products that satisfy their needs. Personalized recommender systems address this issue by offering products that really interest the customer. Currently, there are varieties of recommender systems available in the literature and a number of these systems use Collaborative Filtering(CF) or data mining as an underlying technology for offering recommendations. However, such systems are proven to be replete with well-known problems of sparsity and scalability. In this paper, we present a framework of agent-based personalized recommender system that combines the strengths of collaborative filtering and data mining for offering superior recommendations. INTRODUCTION The advent of electronic commerce and its rapid growth has given a lot of opportunities for customers to choose among a large set of products. However, the customers are often faced with increased difficulty of processing the information overload before they select the products that meet their needs. One way to overcome this problem is to use personalized recommender systems that retrieve the information a consumer desires and help him determine which one to buy (Schafer et al. 2001). E-commerce sites to recommend products to their customers use recommender systems. Currently, there are varieties of recommender systems available in the literature (Cho et al. 2002; Resnick et al. 1994; Shardanand and Maes 1995). A number of these systems use collaborative filtering as an underlying technique for offering personalized recommendations. CF technique primarily matches the similarities and dissimilarities among different customers' tastes and preferences to offer recommendations to the target customer. The technique has been applied in a number of domains such as recommending web pages, movies, articles, and products (Resnick et al. 1994; Shardanand and Maes 1995). However, the technique is replete with problems of sparsity and scalability (claypool et al. 1999; Sarwar et al 2000). Recently researchers have shown that web usage mining can be used effectively to avoid some of the pitfalls of CF like obtaining subjective user ratings or registration based personal preferences. Cho et al. (2002) have suggested a methodology for analyzing the click-stream (visitors' path through a website) data (through web usage mining) to learn customer preferences and product associations. MOTIVATION Significant amount of research has been done in recommending movies, CDs, DVDs, and Usenet newsgroup postings. Such systems typically use explicit ratings of customers to offer recommendations to new customers whose tastes and preferences match with those of existing customers. As CF based methods suffer from issues of sparsity and scalability, (Cho et al. 2002), using a similar approach would not be a suitable strategy for recommending products in domains such as supermarket products, apparels, automobiles etc. Recommender systems that use data mining based methods also suffer from issues of poor quality rules (Prassas et ai. 2001) and scalability (Zaki and Gouda 2001). Given the above limitations of the current recommendation systems, it offers significant scope for research in this fertile field. As more and more firms are conducting their business online, the need for recommander system as an essential business tool has become more of a necessity than a luxury. This scenario gives us immense motivation to venture in this fertile field of research. CONTRIBUTIONS We suggest an agent-based methodology that can offer better recommendations in the online retailing environment. The primary contributions of this paper are: (a) It provides a novel methodology for offering customized recommendations based on which phase of shopping cycle a customer is in (for products that are purchased more frequently), and (b) it offers a methodology for improving the recommendation performance for products that are purchased less frequently, by capitalizing on the customer profiles of products, that are purchased more frequently. …