DASH: Disaster-Aware Social-Semantic Enhanced Hypergraph Convolution Network for Human Mobility Forecasting

Large-scale disasters significantly disrupt human mobility patterns through evacuations, traffic chaos, and population displacement, creating critical challenges for crisis response and urban management. While spatiotemporal graph neural networks have shown success in regular mobility forecasting, they struggle with disaster-induced mobility due to sudden onsets, non-periodicity, and strong spatial locality. Although social media provides valuable disaster-related mobility insights, existing approaches fail to fully exploit semantic information and suffer from temporal information leakage in hypergraph architectures. To address these limitations, we propose Disaster-Aware Social-Semantic Enhanced Hypergraph (DASH), a novel framework for human mobility forecasting under disaster conditions. DASH consists of three core components: (1) a Disaster Semantic Extractor (DSE) that leverages rich semantic cues from social media beyond simple post volumes; (2) a Contrastive Mobility Encoder (CME) that learns discriminative representations capturing discrepancies between various mobility patterns; and (3) a Dynamic Masked Temporal Hypergraph Convolution Network (DMTHGCN) with a novel temporal masking mechanism that prevents future information leakage while modeling high-order spatial relationships. Our tri-modal gating mechanism adaptively fuses mobility, semantic, and volume modalities to enhance disaster sensitivity while suppressing noise. Extensive experiments on three real-world disaster datasets demonstrate that DASH outperforms state-of-the-art baselines, achieving average improvements of 35.44% in RMSE and 37.11% in MAE over all evaluated baselines. Ablation studies confirm the effectiveness of each component, with the contrastive mobility encoder showing the most substantial contribution to performance gains. Moreover, DASH allows potential integration with urban foundation models to inject disaster-aware social sensing into numerical backbones.