Sequential Recommendation with Recurrent Convolutional Model

Abstract

Personalized sequential recommendation refers to making recommendation based on users’ historical consumption behaviors. Most works based on RNN only model long-term patterns, which fail to capture skip behaviors. Contrarily, the CNN-based model whose target is to handle this problem can only leverage part of sequential behaviors and ignores global patterns, which limits its performance. In this paper, we propose a Recurrent Convolutional Recommendation Model (RCRM) to simultaneously catch global and local patterns. Specifically, we employ a recurrent layer to capture global patterns and a convolutional layer to extract local patterns. An attention mechanism is then introduced to generate the final attentive local pattern, which can further concatenate with global patterns to predict next item. We conduct extensive experiments on two benchmark datasets and the results demonstrate that RCRM outperforms state-of-the-art baselines by a large margin over a variety of common evaluation metrics.