Next-generation point of interest (POI) recommendation leveraging users' spatiotemporal movements and social relationships is a research topic in business intelligence. Previous studies have modeled spatial and temporal variations separately, resulting in misaligned representations of the same spatiotemporal core nodes. This leads to unnecessary information generation during the fusion process, increased model uncertainty, and reduced interpretability. To address this issue, this study proposes DiMuST, a socially enhanced POI recommendation model based on separate representation learning from multiple spatiotemporal variation graphs. DiMuST utilizes a novel Disentangled Variational Multiplex Graph Auto-Encoder (DAE) that separates shared and private distributions using a multi-spatiotemporal graph strategy. The DAE fuses shared features through the Product of Experts (PoE) mechanism and removes noise from private features through contrastive constraints. DiMuST effectively captures the spatiotemporal variation representation of POIs while preserving the inherent correlations between spatiotemporal relationships. Experimental results on two challenging datasets demonstrate that DiMuST significantly outperforms existing methods across multiple metrics.
