Wasserstein Adversarial Variational Autoencoder for Sequential Recommendation

Abstract

Variational autoencoders (VAEs) have shown unique advantages as a generative model for sequence recommendation. The core of VAEs is the reconstruction of error targets through similarity metrics to provide a supervised signal for training. However, VAE reconstruction tends to generate non-realistic outputs, which severely affects the accuracy of sequential recommendation. To solve the above problem, in this paper, we propose a new framework called Wasserstein Adversarial Variational Autoencoder (WAVAE) for Sequential Recommendation. In WAVAE, the VAE first combines with the Generative Adversarial Network (GAN) network to differentiate the true and false samples by introducing an adversarially trained discriminative network, and then the learning feature representation in the discriminative network is used as the basis for the VAE reconstruction target so that the VAE tends to generate true samples. We further used Wasserstein loss to optimise the training process, with the aim of avoiding the gradient disappearance problem that occurs when the above-mentioned adversarial network is trained on discrete data, ensuring that the VAE can obtain accurate reconstruction targets through adversarial learning. In addition, we concatenate the original samples with their labels as input to control the generated content and thus control the generated sample tendency. Finally, we conduct experiments on several real datasets to evaluate the model, and the experiment results show that our model outperforms the state-of-the-art baselines significantly.