Pair-wise Preference Relation based Probabilistic Matrix Factorization for Collaborative Filtering in Recommender System

Abstract

Matrix Factorization (MF) is one of the most popular techniques used in Collaborative Filtering (CF) based Recommender System (RS). Most of the MF methods tend to remove sparsity or predict missing ratings. However, the prime objective of any RS is to generate recommendation of items, where items having higher preferences should be placed at higher positions in the recommendation list. Further, the integration of user and item features ( i.e. Side Information) in any RS model, increases the quality of recommendation. Keeping these in mind, we propose a Probabilistic MF (PMF) model that takes Preference Relation as input (instead of ratings) for generating efficient ranking of items. The user and item side information are integrated into the model using matrix co-factorization technique. Also, user and item neighborhood information (known as second-order interaction among similar users or items) are integrated into the model. The use of PMF enables the proposed method to capture user-item interaction of higher-order. Hence, the proposed RS model can capture all the desired user-item interaction information. Experimental results obtained using two ‘MovieLens’ datasets indicate the superiority in recommendation quality of the proposed model over the recent state-of-the-art MF methods.