SHGAE: Social Hypergraph AutoEncoder for Friendship Inference

Yujie Li,Yan Chen,Tianliang Qi,Feng Tian,Yaqiang Wu,Qianying Wang

DOI: https://doi.org/10.1007/978-3-031-44223-0_44

2023-01-01

Abstract:Location-Based Social Networks (LBSNs) present a significant challenge for inferring social relationships from both social networks and user mobility. While traditional rule-based walk graph representation learning methods predict friendship based on user proximity, they fail to distinguish contributions of different mobile semantics (temporal, spatial, and activity semantics). On the other hand, graph-based autoencoder models have shown promising results, but they are not suitable for heterogeneous information in LBSNs, and they perform poorly when users lack initial features. In this paper, we propose the Social Hypergraph Autoencoder (SHGAE) model, a novel autoencoder designed specifically for social hypergraphs formed by LBSNs data, which combines the strengths of these two methods. We initialize nodes vectors via hypergraph-jump-walk embedding strategy to capture features of the hypergraph, then use a well-designed autoencoder with heterogeneous message passing and attention mechanisms to model different semantic node influences. Extensive experiments demonstrate that our model outperforms state-of-the-art methods on the social relationship inference task. Moreover, in the ablation study, we find that our two proposed modules contribute differently to datasets with different sparsity, which can provide valuable insights for future research.