Weighted Hybrid Model Based Product Recommender System using RBM and Matrix Factorization

Abstract

Collaborative filtering is one of the most widely used techniques in the recommendation system. Its major challenge is how to handle sparseness of the data. It performs better with enough user rating history records but behaves badly in sparseness of the data. RBM is a generative model act as a foundation block of deep learning that can handle sparseness of the data to a great extent. However, rating prediction by a single model is not preferable. Hybrid model by combining the results of more than one models yield better results. In this paper, we have performed an empirical analysis of weighted hybrid CF method by combining RBM with other matrix factorization model such as SVD, SVD++, NMF and Random recommendation model. Product recommendation can be seen as matrix completion problem, as a user’s preference for different products can be represented in a sparse matrix. Missing ratings can be predicted by the ensemble learning models. In the proposed framework, apparel items are recommended to the active users by combining the results of different models. Effectiveness of the models has been tested with the Amazon fashion rating data.