Mining Latent Attributes in Neighborhood for Recommender Systems

Abstract

Neighborhood-based collaborative filtering (CF) algorithms have been extensively studied and discussed. In the traditional way of these methods, user-based CF predicts a target user’s preference for an item based on the integrated preference of the user’s neighbors for the item, and item-based CF is based on the integrated preference of the user’s preference for the item’s neighbors. Both the two ways underestimate the effect of structure of the target user or item’s neighbors. That is, for instance, these neighbors may form two distinct groups: some neighbors like the target item or give high ratings; on the other hand, some neighbors dislike the target item or give low ratings. The difference between the two groups may influence user’s choice. As an extension of neighborhood-based collaborative filtering, this paper focuses on the analysis of such structure by mining latent attributes of users or items’ neighborhood, and corresponding correlations with users’ preference by several popular data mining techniques. Mining latent attributes and experiment evaluation were conducted on MovieLens dataset. The experimental results reveal that the proposed method can improve the performance of pure user-based and item-based collaborative filtering algorithm.