Incremental collaborative filtering based on Mahalanobis distance and fuzzy membership for recommender systems

Abstract

Recommender systems, as an effective personalization approach, can suggest best-suited items (products or services) to particular users based on their explicit and implicit preferences by applying information filtering technology. Collaborative filtering (CF) method is currently the most popular and widely adopted recommendation approach. It works by collecting user ratings for items in a given domain and by computing the similarity between the profiles of several users in order to recommend items. Current similarity measures and models updated by traditional model-based CF have, however, shortcomings with respect to accuracy of prediction and scalability of recommender systems. To overcome these problems, here an incremental CF algorithm based on the Mahalanobis distance is presented. The algorithm has two phases: the learning phase, in which models of similar users are constructed incrementally, and the prediction phase, in which interested users are clustered by measuring their similarity to existing clusters in a model. To handle confusion of decision making on overlapping clusters, fuzzy sets are employed, and the degree of membership to them is expressed by the Mahalanobis radial basis function. Experimental results demonstrate that the proposed algorithm leads to improved prediction accuracy and prevents the scalability problem in recommendation systems.

Keywords:

• recommender systems
• collaborative filtering
• learning algorithms
• Mahalanobis distance
• fuzzy sets