Abstract:
The proliferation of smart mobile devices and social networks has streamlined location sharing, enabling the analysis of check-in data for predicting human movement and enhancing point of interest (POI) recommendations. This predictive capability is pivotal for various intelligent location-based services, including crowd flow prediction and business recommendations. However, sparse user-POI interactions due to privacy concerns pose challenges to accurate prediction. To tackle these challenges, we propose a location prediction framework that integrates temporal-spatial dependencies and sequential influences. While traditional methods like matrix Factorization and Markov Chains have limitations in capturing complex behavioral patterns, recent advancements in deep learning, particularly Recurrent Neural Networks (RNNs) and Graph Convolutional Networks (GCNs) and knowledge graphs, offer promising avenues for improved prediction accuracy. Our proposed model, PULSE (Predictive User Location and Spatiotemporal Experience), leverages graph neural networks to capture both long-term life patterns and short-term behavioral preferences from user trajectories. By considering spatial and temporal dependencies separately and fusing them effectively, PULSE demonstrates superior performance compared to existing approaches, as validated through extensive experiments on real-world datasets. In summary, PULSE represents a novel approach to location prediction that harnesses the power of deep learning and Knowledge networks / Graphs, offering a comprehensive solution to address the complexities of human mobility forecasting and preferences history.
