Sequential Behavior Modeling for Next Micro-Video Recommendation with Collaborative Transformer

Micro-video recommendation is important for micro-video social platform, which provides its users with micro-videos they may interested in. In this paper, we propose a new variant of Transformer, alleviating the drawbacks of Recurrent Neural Networks (RNN) which tend to compress all history records in a fixed hidden representation, to model users' sequential behavior for next micro-video recommendation. Firstly, we employ self-attention to capture micro-video's multi-modal features of different importance. Secondly, we make use of multi-head attention to learn users' preference from historical records. We show how the Transformer combined with Collaborative Filtering (CF) and user-video sequential interaction can be used to perform next micro-video recommendation. Extensive experiments are conducted on two collected micro-video datasets, i.e., Toffee and TikTok. The experimental results demonstrate the proposed method is more effective compared with several state-of-the-art sequential recommendation methods.