Session-based recommendation via flow-based deep generative networks and Bayesian inference

Abstract

We present a novel generative Session-Based Recommendation (SBR) framework, called VAriational SEssion-based Recommendation (VASER) – a non-linear probabilistic methodology allowing Bayesian inference for flexible parameter estimation of sequential recommendations. Instead of directly applying extended Variational AutoEncoders (VAE) to SBR, the proposed method introduces normalizing flows to estimate the probabilistic posterior, which is more effective than the agnostic presumed prior approximation used in existing deep generative recommendation approaches. We also combine the effectiveness of both stochastic and amortized variational inference to reduce the inference gaps and to alleviate the underfitting problem of variational recommendation. We propose two specific implementations of VASER, both of which explore soft attention mechanism to upweight the important clicks in a session and show that one of them, treating the attention vector as an auxiliary latent factor, can make the variational distribution more expressive, and thus improves the recommendation accuracy over the widely used deterministic attention approaches. Empirically, we show that the proposed models significantly outperform several state-of-the-art baselines, including the recently-proposed RNN/VAE-based approaches, on several real-world datasets.