Self-Attention Network for Session-Based Recommendation With Streaming Data Input

In the current era of the rapid development of big data, it has become increasingly critical and practical to study recommender systems with streaming data input. However, the recommender system is often faced with the challenge that the history records of new users or anonymous users are not available. Specifically, session-based recommendation, which aims to predict a user's next actions, is a typical task to overcome the challenge. To capture a user's long-term preference in session-based recommendations, recurrent neural networks (RNN)-based models have been widely applied with impressive results, but the inherent sequential nature of RNNs prevents parallelism within training examples, which is critical in long sessions because memory constraints limit batching across instances. In this paper, we propose a novel method, i.e., self-attention network for session-based recommendation (SANSR), which is based on only attention mechanisms, dispensing with recurrence, and supports parallelism in the session. The proposed model attempts to find items that are relevant based on previous time steps in the ongoing session and to assign them different weights to predict the next item. The extensive experiments are conducted on two real-world datasets, and the experimental results show that our proposed model is superior to the state-of-the-art methods.