Abstract:

The problem of user-POI prediction, based on the user’s historical check-in records, determines whether a user checks in a specific POI. However, the user-POI data has a long-tail distribution phenomenon. To address this data sparsity challenge, existing work disentangled the geographical neighbor effect and the geographical sequence effect through self-supervised learning to improve the interpretability and accuracy of the POI prediction task. This paper further introduced the semantic sequence effect, and proposed an improved disentangled graph embedding model. The model used the pairwise constraints of point-of-interests in the geographic space and semantic space, and was based on the feature expression, feature modification, feature decoupling and multi-layer perceptron fusion of the influencing factors in the geographic coordinate space and the semantic category space. The geographic level could be combined with the semantic level to better predict the user’s access to the POI. Experimental results show that the proposed method can still achieve good prediction effects on sparse datasets.

Key words: point of interest prediction, self-supervised disentanglement, graph embedding, joint prediction