Deep hybrid collaborative filtering for E-commerce product recommendation system

Abstract

Deep learning based approaches for collaborative filtering have proven highly successful in modern recommendation system research. In this work, we explore Variational Autoencoders to build a product recommendation system. We utilize modern frameworks of probabilistic deep learning to carry out an empirical analysis on real world datasets and compare it with traditional matrix factorization methods.

Matrix factorization methods for collaborative filtering suffered from sparsity in the data and under performed on scalability measures. We use these methods as baselines and strive to outperform them. We augment a standard Variational Autoencoder with a regularization parameter that partially anneals the divergence term in the objective function and use multinomial likelihood to model the user-item interaction data.

To analyse our methods, we use two recently collected datasets comprising of user and item interactions. We run various experiments on these datasets to find the best possible model. We report a quantitative comparison between the baselines and our approach on various metrics. We reach satisfactory results suggesting that the partially regularized Variational Autoencoder framework with a multinomial likelihood is well suited for the collaborative filtering task on real world data.