融合时空网络和自注意力的兴趣点序列推荐

Sequential recommendation of point of interest combining spatio-temporal network with self-attention

针对现有的序列推荐算法仅利用长短期顺序行为和用户交互行为进行推荐,没有充分考虑用户交互行为之间的时空间隔信息对用户推荐序列更深层影响的问题,提出一种融合时空网络和自注意力的兴趣点序列推荐模型。将用户签到之间的时间和空间间隔信息融入门控循环单元网络,使用用户的历史签到序列信息获取用户的偏好,通过自注意力机制对签到地点进行建模,获得用户对于模型的权重序列,通过签到地点与候选地点的时间间隔和空间间隔匹配兴趣点,为用户推荐一个兴趣点序列。在两个数据集上的实验结果表明,提出方法在召回率上优于之前先进的方法。

To solve the problem that the existing sequential recommendation algorithms only use long-term and short-term sequential behaviors and user interaction behaviors for recommendation, but fail to fully consider the deeper influence of the spatio-temporal interval information between user interaction behaviors on the user recommendation sequence, a sequential recommendation model of point of interest combining spatio-temporal network with self-attention was proposed. The spatiotemporal interval information between user check-ins was integrated into the gated recurrent unit network and the user’s historical check-in sequential information was used to obtain the user’s preferences. The check-in location was modeled through the self-attention mechanism to obtain the user’s model. The point of interest was matched with the time interval and space interval between the check-in location and the candidate location to recommend a sequence of points of interest for the user. Experimental results show that the proposed method is superior to the previous advanced methods in recall rate on two datasets.