SeMob: Semantic Synthesis for Dynamic Urban Mobility Prediction

TL;DR

SeMob addresses the challenge of predicting urban mobility under event-driven disruption by introducing an LLM-powered semantic synthesis pipeline with a multi-agent information-extraction framework. It couples a two-stage progressive fusion of textual context with spatiotemporal signals to produce context-aware forecasts, showing substantial improvements over traditional spatiotemporal models, especially near event venues and times. Key contributions include the automated extraction and reasoning of event-related text, a context-enriched dataset, and a fusion mechanism that dynamically weighs multimodal inputs. The approach demonstrates practical impact for minute-level urban forecasting while highlighting the value and ethical considerations of integrating unstructured textual data into mobility prediction.

Abstract

Human mobility prediction is vital for urban services, but often fails to account for abrupt changes from external events. Existing spatiotemporal models struggle to leverage textual descriptions detailing these events. We propose SeMob, an LLM-powered semantic synthesis pipeline for dynamic mobility prediction. Specifically, SeMob employs a multi-agent framework where LLM-based agents automatically extract and reason about spatiotemporally related text from complex online texts. Fine-grained relevant contexts are then incorporated with spatiotemporal data through our proposed innovative progressive fusion architecture. The rich pre-trained event prior contributes enriched insights about event-driven prediction, and hence results in a more aligned forecasting model. Evaluated on a dataset constructed through our pipeline, SeMob achieves maximal reductions of 13.92% in MAE and 11.12% in RMSE compared to the spatiotemporal model. Notably, the framework exhibits pronounced superiority especially within spatiotemporal regions close to an event's location and time of occurrence.

Figures (7)

• Figure 1: Challenges in event-based mobility prediction. (a) Spatiotemporal models, effective during regular days, can exhibit higher errors than the simpler LSTM. (b) Complex event semantics drive pattern variance, even in similar event types.
• Figure 2: SeMob framework. Multi-agents iteratively collect spatiotemporally relevant texts for TST multimodal prediction and refine the text filtering logic using prediction feedback.
• Figure 3: TST architecture. We first (1) synthesizes dynamic event signatures by fusing textual embeddings with evolving temporal contexts and (2) integrates spatiotemporal data with these signatures for fine-grained, context-aware mobility predictions.
• Figure 4: Comparative visualization of MAE reduction derived from basic event information, public reaction, and inferred traffic conditions versus spatiotemporal information alone.
• Figure 5: Performance of different time windows across training data sizes. ‘Base’ indicates the performance of GWNET without event information.