Heterogeneous spatio-temporal graph contrastive learning for point-of-interest recommendation

As one of the most crucial topics in the recommendation system field, Point-of-Interest (POI) recommendation aims to recommending potential interesting POIs to users. Recently, graph neural networks have been successfully used to model interaction and spatio-temporal information in POI recommendations, but the data sparsity of POI recommendations affects the training of GNNs. Although current GNN-based methods show their superior performance, this paper argues that they are flawed in two aspects: The first one is coarse granularity for modelling heterogeneity in POI recommendation data. Existing approaches tend to construct only a bipartite graph of users and POIs to capture their node heterogeneity and simple interaction between them. Nevertheless, due to the existence of time and space factors, there are more kind of heterogeneities and more complex relationships in POI recommendation, which is often overlooked by existing GNN-based methods. To deal with the wide variety of heterogeneity, some work (e.g., STGCN) constitute a complex multi-layer graph to cover various heterogeneities, but the newest work (i.e., MPGRec) argues that this may reduce the recommendation performance because it mixes too much information while message passing and introduces additional noise. Therefore, how to reduce the impact of noise on GNN training while using various heterogeneous information is still a key topic that needs to be explored. Another shortcoming of existing works is insufficient consideration of interaction sparsity issues. Recommender systems often suffer from data sparsity, making GNN-based models challenging to learn high-quality node representations or susceptible to interaction noise. In other recommendation scenarios, some recent works (e.g., SGL, NCL) have focused on reducing the impact of data sparsity problems on GNN models. However, for POI recommendations, this issue still receives little attention. To address these two limitations, we propose a novel POI recommendation method, named HestGCL). To model the heterogeneity in POI recommendation scenarios at a finer granularity, we build a heterogeneous spatio-temporal graph including three kinds of nodes (i.e., User, POI and Location) and three kinds of relations, which helps uncover the influence of heterogeneous information on recommendations. To address the challenges posed by data sparsity and spatio-temporal noises to GNN models, inspired by self-supervised learning, this paper proposes a cross-view contrastive learning technique for spatio-temporal heterogeneous graphs. Specifically, HestGCL first splits the complete heterogeneous graph into a spatial view and a temporal view. Then, HestGCL designs spatial-aware and temporal-aware graph neural networks for spatial and temporal views, respectively. Finally, HestGCL uses the node representations obtained from each view for contrastive learning. Experimental results on three public datasets demonstrate that our HestGCL model achieves consistent and significant improvement over state-of-the-art baseline methods on the POI recommendation task. The relative improvements of Recall@50 are 8.83\%, 14.61\%, and 6.86\% on Foursquare, Gowalla and Meituan, respectively. Ablation studies and hyper-parameter experiments further demonstrate the effectiveness and robustness of HestGCL.

[1] J. Liu et al., "Heterogeneous Spatio-Temporal Graph Contrastive Learning for Point-of-Interest Recommendation," in Tsinghua Science and Technology, vol. 30, no. 1, pp. 186-197, February 2025, doi: 10.26599/TST.2023.9010148.

About Tsinghua Science and Technology

Tsinghua Science and Technology is sponsored by Tsinghua University and published bimonthly, 2023 Impact Factor of 5.2, ranking in Q1 in the "Computer Science, Software Engineering", "Computer Science, Information System", and "Engineering, Electrical & Electronic" areas in SCIE, according to JCR 2023. This journal aims at presenting the achievements in computer science, electronic engineering, and other IT fields. This journal has been indexed by SCIE, EI, Scopus, etc. Contributions all over the world are welcome.