Learning User Embeddings from LBSN Augmented by Spatial-Temporal Constraints

Location-Based Social Networks (LBSN) have generated massive users’ check-in data, which enables various services such as friend recommendation. Due to the recent success of Heterogeneous Graph Neural Networks (HGNN), LBSNs are modeled as heterogeneous graphs and analyzed by HGNNs. However, most existing methods either follow a supervised/semi-supervised learning, or ignore the spatial-temporal information in LBSN. To address the above issues, in this paper we propose a novel model User Embedding with Spatial-Temporal Constraints (UE-STC) to learn user embeddings for friend recommendation. Specifically, the UE-STC model first augment the LBSN graph by exploiting spatial constraints and temporal constraints, which alleviates both the sparsity issue and the noise issue. Then, the UE-STC model learn user embeddings by contrastive learning on views designed from two aspects and two levels by exploiting semantic and heterogeneous information carried by meta-paths, which captures both the consistency and the complementarity of information. The experiment results on three real-world LBSN datasets show the UE-STC model outperforms state-of-the-art baselines for friend recommendation, verifying the effectiveness of the UE-STC model in learning user embeddings form LBSN.