JoVA-hinge: joint variational autoencoders for personalized recommendation with implicit feedback

Abstract

Recently, Variational Autoencoders (VAEs) have shown remarkable performance in collaborative filtering (CF) with implicit feedback. These existing recommendation models learn user representations to reconstruct or predict user preferences. However, existing VAE-based recommendation models learn user and item representations separately. This thesis introduces joint variational autoencoders (JoVA). JoVA, as an ensemble of two VAEs, simultaneously and jointly learns both user-user and item-item correlations and collectively reconstructs and predicts user preferences. Moreover, a variant of JoVA, referred to as JoVA-Hinge, is introduced to improve recommendation quality. JoVA-Hinge incorporates pairwise ranking loss to VAE's losses. Extensive experiments on multiple real-world datasets show that our model can outperform state-of-the-art under a variety of commonly-used metrics. Our empirical experiments also confirm that JoVA-Hinge offers better results than existing methods for cold-start users with limited training data.