A Median Concept for Model-Based Collaborative Filtering

Abstract

Model-based collaborative filtering has widely been used to allow the recommender systems to learn to recognize complex patterns of user preferences. So far many clustering models have been investigated to solve the shortcomings of memory-based algorithms. Mostly, the user preferences are in form of sparse matrix which comprises of many noises. Predictions of recommendations for any active users are unavoidably effected by outlier data. While the previous works rely on solving general clustering problems, we propose a general framework to apply a concept of generalized median to collaborative filtering. From a general point of view, given a distance function d(p, q), the essential information of a given set of patterns S in arbitrary space U is covered by a generalized median p ̅∈U that minimizes the sum of distances to all patterns from S. This concept has found various applications in dealing with strings, graphs, curves, and clusterings. In our current work pseudo user rating matrix which obtains by item-based collaborative filtering to control a degree of data sparsity. The various model-based collaborative filtering by generalized median concept is applied to compare overall performance in reducing the leftover effects of outliers. Additionally, we compare our approach to various collaborative filtering algorithms for performance evaluation.