STAC-HNN: a spatio-temporal auto-correlation based location recommendation using hypergraph neural network

Location recommendation normally used recurrent neural networks (RNNs), which often overlooked higher-order dependencies and similar subsequences that encapsulate latent user preferences. Meanwhile category information was underutilized. In this paper, we propose STAC-HNN, a spatio-temporal auto-correlation based model using hypergraph neural networks, to address these limitations. STAC-HNN captures both local and global signal representations: the global module leverages hypergraph neural networks to model high-order interactions among user trajectories and POI categories, while the local module employs a spatio-temporal weight matrix and self-attention mechanisms to uncover relations between non-adjacent locations and visits. Additionally, a bi-task network independently and collaboratively learns users’ POI and category preferences, enhancing model performance. Experiments on three real-world datasets show that STAC-HNN significantly outperforms state-of-the-art methods in next POI recommendation. To facilitate future research, we release the codes at https://github.com/chenshone/STAC-HNN.