Analysis of customer behavior patterns for enhancing the collaborative recommendations

Abstract

Recommender systems are personalized information filtering techniques to help customers find the product they would like to purchase. These systems are achieving a widespread success in online marketplace. Among them, collaborative recommendation has been known to be the most successful recommendation technology. Collaborative recommendation system basically requires a customer profile to preserve the ratings about the preference of customers. The tremendous growth of product and customers nowadays makes it difficult for the systems acquire the direct ratings of preference from customers. Instead, most systems commonly employ implicit ratings in which customers’ preference is gathered without customer’s intervention. However, applying implicit ratings to collaborative recommendation poses some research issues that must be addressed. The first issue is the rating scale problem. As an alternative for ratings scales, cardinal scale is the more accurate scale type due to its high sensitivity. It has been widely used with explicit ratings helping customers to represent their preference closely to the actual things. However, when used in implicit ratings, cardinal scale may increase the estimation error by increasing the variance of estimated values. Therefore, the use of ordinal scale should be considered as an alternative for ratings scale in implicit ratings. The second issue is preference compromise problem. In implicit ratings, customer’s preference is often collected by analyzing the behavior histories recorded on the Web log data. A well-known approach for implicit ratings, Web usage mining (WUM) discovers customer behavior patterns from the web log and as a result, a great amount of preference information is collected. As individual information collected is partial and sometimes conflictive each other, they should be aggregated to become a complete preference of customers over all the items. Further, the preference should be represented with ordinal sca...