This paper proposes a multi-scale spatial-temporal decoupled predictor (MSTDP), a medium- to long-term prediction model for individual mobility patterns. Unlike previous studies that primarily focus on predicting next locations, our approach aims to predict daily or weekly travel paths, considering applications that require long-term predictions, such as traffic management and epidemic management. MSTDP efficiently extracts spatial and temporal information by decomposing daily travel paths into location-duration chains. It uses a hierarchical encoder to model multi-scale temporal patterns, such as daily repetition and weekly periodicity, and a transformer-based decoder to globally focus on prediction information within the location or duration chains. Furthermore, a spatial heterogeneous graph learner is introduced to capture multi-scale spatial relationships, enhancing semantic representations. Extensive experiments on large-scale mobile phone records from five cities, including Boston, Los Angeles, the San Francisco Bay Area, Shanghai, and Tokyo, demonstrate the advantages of MSTDP. Applying it to an epidemic model in Boston, our model achieves a remarkable 62.8% reduction in the MAE of the cumulative number of new cases compared to the existing best-performing baseline model.
