A hierarchical contextual attention-based network for sequential recommendation

Qiang Cui, Shu Wu, Yan Huang, Liang Wang

Abstract: The sequential recommendation is one of the most fundamental tasks for Web applications. Recently, recurrent neural network (RNN) based methods become popular and show effectiveness in many sequential recommendation tasks, such as next-basket recommendation and location prediction. The last hidden state of RNN is usually applied as the sequence’s representation to make recommendations. RNN can capture the long-term interest with the help of gated activations or regularizers but has difficulty in acquiring the short-term interest due to the ordered modeling. In this work, we aim to strengthen the short-term interest, because it is beneficial to generate responsive recommendation according to recent behaviors. Accordingly, we propose a Hierarchical Contextual Attention-based (HCA) network. First, RNN is extended to model several adjacent factors at each time step. Such kind of multiple factors can be considered as a context where the short-term interest comes from. Then, within the context, the attention mechanism is used to find the important items that contribute to the short-term interest. This contextual attention-based technique is conducted on the input and hidden state of RNN respectively. In this way, we can relieve the limitation of ordered modeling of RNN, model the complicated correlations among recent factors, and strengthen the short-term interest. Experiments on two real-world datasets show that HCA can effectively generate the personalized ranking list and achieve considerable improvements.