Individual weekly activity patterns are essential for enhancing activity-based mobility models, yet existing methods mainly focus on single-day behaviors, while overlooking temporal dependencies across days, and struggle to balance predictive accuracy with interpretability. To address the gap, this study proposes AP-WASG, a three-layer, activity-pattern-based weekly activity sequence generation framework grounded in natural language processing. The proposed method first casts an individual weekly timeline as text, leveraging latent Dirichlet allocation (LDA) to distill complex sequences into a concise topic space and then grouping them into seven activity patterns. Secondly, an XGBoost classifier forges a clear link between these patterns and individuals’ socio-demographics and built environment, making it possible to predict one’s dominant activity pattern in new contexts. A Transformer-based architecture is then proposed to generate individual weekly activity sequences with the socio-demographics-enhanced texts, achieving cosine similarity scores of 0.74–0.92 with real-world data, representing 12–33% improvement over pattern-unconstrained approaches. The hybrid framework addresses traditional limitations in big data applications while maintaining interpretability through pattern constraints. The results provide practical support for personalized urban planning, transportation management, and policy analysis, including cold-start and scenario applications in unseen neighborhoods and planning years.
原文链接
分享
引用
