A Novel Top-N Recommendation Approach Based on Conditional Variational Auto-Encoder

Abstract

Personalized recommendation has continuously received attention due to its great commercial value in business. Recently variational auto-encoder is employed in top-N recommendation for its effectiveness in deep collaborative filtering. The key challenge of model-based collaborative filtering is to develop effective latent factors representations with user-item interaction records. In this paper, we present a new class of conditional variational auto-encoders (CVAEs) that utilizes the fact of similar users tending to associate with each other on purchasing preference. This type of conditional variational auto-encoder concentrates on learning with label verification signals to ensure an exclusive latent mean factor for users with the same labels. Moreover, to handle complex multi-label combinations, we extend the model with a split-merge framework by learning labels of different conditional attributes separately and then merge the results from multiple prediction pools. Extensive experiments are conducted on two real-life datasets to simulate both user-based and item-based recommendation scenarios. Experimental results are favorable when comparing with the state-of-art methods.