Context-aware seq2seq translation model for sequential recommendation

Abstract

Abstract Context information, such as product category, plays a vital role in sequential recommendations. Recently, there has been a growing interest in context-aware sequential recommender systems. However, in previous studies, contexts have often been treated as auxiliary information without the consideration of the inter-sequence dependency between the item sequence and the context sequence. Such a dependency provides valuable details for predicting a user’s future behavior. For example, a user may buy electronic accessories after buying an electronic product. In this paper, we propose a context-aware seq2seq translation model to capture the inter-sequence dependency for sequential recommendations. The key component in our model is a tripled seq2seq translation architecture with an injected variational autoencoder (VAE). The tripled architecture, consisting of forward and backward translation, naturally encodes bi-directional inter-sequence dependency. Moreover, the injected VAE enables the translation process to redress the semantic imbalance between context and item. We conduct extensive experiments on four real-world datasets. The results show the superior performance of our model over the state-of-the-art baselines. The code and datasets are available at https://github.com/NLPWM-WHU/CAST.