Future locations prediction with multi-graph attention networks based on spatial–temporal LSTM framework

• Published in: The Journal of supercomputing Vol. 80; no. 14; pp. 20020 - 20041
• Main Authors: Li, Zhao-Yang, Shao, Xin-Hui
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.09.2024; Springer Nature B.V
• Subjects: Attention; Compilers; Computer Science; Interpreters; Location based services; Processor Architectures; Programming Languages; Spatial data; Spatiotemporal data; Traffic; Location prediction; Spatial–temporal dependency; Graph learning; Seq2Seq framework
• ISSN: 0920-8542; 1573-0484
• DOI: 10.1007/s11227-024-06249-9

Studies on human mobility from abundant trajectory data have become more and more popular with the development of location-based services. Prediction for locations people may visit in the future is a significant task, helping to make visiting recommendations and manage traffic conditions. Different from other time series prediction tasks, location prediction is temporally dependent as well as spatial-aware. In this paper, we propose a novel multi-graph attention network with sequence-to-sequence structures based on spatial–temporal long short-term memory to predict future locations. Specifically, we build three graphs with movements in geographic space and apply graph attention networks to explore the latent spatial associations among geographic regions. Additionally, we come up with spatial–temporal long short-term memory and use it to establish a sequence-to-sequence framework, which collects the temporal dependence as well as some spatial information from history trajectories. The predictions of future location are finally made by aggregating spatial–temporal contexts.