带有时间预测辅助任务的会话式序列推荐

Session-based sequential recommendation aims to predict a target user's behavior in the near future based on its anonymous short-term behavior sequence in a given current session. Since this task setting provides a good chance of considering user dynamic preference, it has gained much attention from both the academic and industrial domains. Most of the existing relevant methods in the literature are attributed to the category of only focusing on predicting the next items to interact for a considered user in a single-task mode, which inevitably overlooks the additional useful semantic information that is contained in the auxiliary task of predicting user behavior time. For general sequential recommendation problems that take events or locations as the input, there exist a few approaches that investigate the ways of predicting the next item and its corresponding time at the same time in a parallel fashion. However, this paradigm is not very consistent with the actual situation of sequential behaviors, where users tend to first determine the intention of an interaction and then choose the appropriate time in the future to perform the interaction. To mitigate the effect of the above issue, this paper develops a novel session-based sequential recommendation method by utilizing the framework of sequential multi-task learning. The devised method possesses two aspects of innovations. First of all, it predicts the interacted items in the next time and further take them as the input of the next time prediction method, thereby this manner empowers the two prediction tasks with sequential dependency. As a result, compared with the parallel fashion of the joint prediction of items and their corresponding timestamps, the proposed model can couple the two tasks more tightly. Secondly, an improved bidirectional time interval aware Self-attention approach is developed in this paper, which enables the item at each position of a target session to have the sequential context information, i.e., the other interacted items and the time intervals between the current item and the other items, from both the left and right sides of the same session. As such, compared with the conventional session-based sequential recommendation approaches that model each session in a one-way mode, this proposed bidirectional approach is welcomed for enhancing the ability of modeling the contextual information of sessions, which in turn obtains more precise user interest representations for later prediction. The comprehensive experiments in this paper are conducted on three publicly available datasets with different origins, including the Tianchi dataset for the e-commerce scenario, the Lastfm dataset used in music listening, and the Foursquare dataset that is generated by spatio-temporal trajectories. The experimental results on the three datasets reveal that: (1)The proposed method improves the adopted strong baselines consistently in terms of different evaluation metrics commonly used in personalized sequential recommendation. In particular, it outperforms the best baseline method, i.e., TiSASRec, on average by 13.51% in terms of the NDCG@5 evaluation metric. (2)Both the sequential multi-task learning part and bidirectional time interval aware Self-attention mechanism can bring positive improvements over the next item and time prediction performance.