Abstract: How to capture the dynamic changes and dependencies of user behavior is a vital issue existing in point-of-interest (POI) recommendation. It mainly faces challenges including data scarcity, difficulty in extracting spatio-temporal sequence features and in capturing users?? individuated differences. In order to address these challenges, this paper proposes a time-sequence POI recommendation method based on spatio-temporal vicinity perception and implicit changes of users?? state. This method is aimed at converting the learning of user behavior into the learning of users?? latent state, combined with distance information to introduce spatial information, which effectively captures users?? mobile characteristics. Firstly, the variational autoencoder is utilized to represent the potential state of users. And then the dependency among the latent states is learnt through the graph neural network so as to capture the time-sequence dependence of user behavior. Furthermore, this paper makes use of the attention mechanism and radial basis function to capture the spatial dependence between the user and location candidate sets. Next, this paper evaluates the frequencies of user visiting each location, hence achieving point-of-interest recommendation. Experimental comparison and analysis on three real datasets demonstrate that the temporal recommendation performance of the proposed method is superior to existing benchmark algorithms.

Key words: point-of-interest (POI) recommendation, variational auto-encoder, graph neural network, attention mechanism

摘要： 如何捕捉用户行为的动态变化和依赖关系是当前兴趣点推荐的一个重要问题，主要面临着数据稀疏、时空序列特征提取难以及用户个性化差异不易捕捉等挑战。为了解决这些挑战，提出了一种基于时空邻域感知及隐含状态变化的时序兴趣点推荐方法。该方法将用户行为的学习转换成了潜在状态的学习，并以一种结合距离信息的方式引入空间信息，有效地捕捉了用户的移动特征。首先，利用变分自编码器表征用户的潜在状态，再通过图神经网络学习到潜在状态之间的依赖关系，从而捕捉到用户行为的时序依赖；然后，利用注意力机制和径向基函数来捕捉用户与地点候选集之间的空间依赖，进而评估用户访问每个地点的概率，实现兴趣点推荐。在三个真实数据集上进行了实验比较和分析，显示了该方法相比于现有的基准算法具有更好的时序推荐性能。

关键词: 兴趣点推荐, 变分自编码器, 图神经网络, 注意力机制