Correlated Wasserstein Autoencoder for Implicit Data Recommendation

Recommender systems for implicit data, e.g., browsing data, have attracted more and more research efforts. Most existing approaches assume the implicit data is i.i.d. which ignores the fact that the real-world data is generally correlated with each other. To cope with this issue, this paper proposes the correlated Wasserstein autoencoders (CWAEs) model to capture data correlation to enhance recommendation peformance. Particularly in the proposed approach, we first formulate correlated data via an undirected acyclic graph and then generalize the undirected acyclic graph to an acyclic graph by averaging all its' maximum acyclic subgraphs. To further enhance model performance, we introduce negative sampling strategy. Experiments are evaluated on Epinions dataset. The widely adopted evaluation criteria, i.e., CRR and NCRR, are adopted to evaluate both baseline models and our proposed approach. Experimental results have demonstrated the superiority of the proposed models.