Point-of-Interest (POI) recommendation is crucial in the recommendation system field. Graph neural networks are used for POI recommendations, but data sparsity affects GNN training. Existing GNN-based methods have two flaws. Firstly, they have coarse granularity for modelling heterogeneity, overlooking complex relationships due to time and space factors. Although some work constructs complex graphs, it may reduce performance by introducing noise. Secondly, they insufficiently consider interaction sparsity issues, with little attention in POI recommendations. To solve these problems, a novel method HestGCL is proposed. It builds a heterogeneous spatio-temporal graph with three node types and three relations to model heterogeneity at a finer granularity. Inspired by self-supervised learning, it uses a cross-view contrastive learning technique, splitting the graph into spatial and temporal views, designing specific graph neural networks, and using node representations for contrastive learning. Experiments on three datasets show that HestGCL outperforms state-of-the-art methods, with relative improvements in Recall@50, and ablation studies prove its effectiveness and robustness.

Researchers from Chinese Academy of Sciences and Peking University introduce DFFPA—a novel method that enhances detection capabilities for new object classes with limited data. DFFPA leverages dual-domain feature fusion and patch-level attention to achieve superior performance. This breakthrough holds good potential for applications in autonomous driving and robotics.