Collaborative Filtering Enhanced by Hybrid Variational Autoencoder Framework

Abstract

Personalized recommendations have become crucial in the current era where nearly every industry has an online existence and users engage in online marketplaces. Historically, collaborative filtering had been addressed utilizing Matrix Factorization, which is a linear method. We build on the work described in reference [11] by presenting a hybrid multi-modal approach for collaborative filtering with implicit feedback that uses VAE (Variational Autoencoders). To improve movie recommendation, we combine user ratings from the Movielens 20M dataset with movie embeddings obtained from a related VAE network. We demonstrate how the network of VAE benefits from including movie embeddings through empirical evidence. We cluster the latent representations of movie and user embeddings attained from a VAE and visualize them.