ReMoE: An Interpretable Mixture-of-Experts Framework for Human Mobility Recovery Forecasting Across Disasters
Organizations: School of AI and Liberal Arts, Beijing Normal-Hong Kong Baptist University, Zhuhai, China · Shenzhen University of Advanced Technology, Shenzhen, China · Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China · Department of Mechanical Engineering, City University of Hong Kong, Hong Kong, China
Abstract
Human mobility often declines sharply following disasters and subsequently exhibits complex recovery dynamics. Effective forecasting of this recovery is critical for timely assessment and resource allocation. However, existing approaches are often developed for individual events and operate as black boxes, limiting their transferability and interpretability beyond the events they were trained on. Recent studies have identified recurring post-disaster mobility recovery types across disasters, suggesting that recovery knowledge may be transferable. Motivated by this, we propose ReMoE, an interpretable mixture-of-experts framework for cross-disaster mobility recovery forecasting. ReMoE uses a feature-guided router to adaptively combine specialized experts that learn these heterogeneous yet reusable recovery types, producing a region-specific recovery pattern. It further refines this pattern through residual correction informed by forecasts of future hazard conditions. We evaluate ReMoE on a large-scale mobility dataset covering 18 disaster events spanning hurricanes, floods, winter storms, and wildfires. ReMoE is competitive in-distribution and achieves the strongest performance on unseen events and hazard types, with larger gains under greater distribution shifts. Further analyses show that ReMoE yields interpretable learning outcomes and maintains robust performance under uncertainty. Broadly, ReMoE offers a transferable and traceable forecaster that can support anticipatory post-disaster assessment.
Figures & tables
| Setting | In-distribution | LOEO | LOHTO | ||||||||||||
| Metric group | Prediction accuracy | Recovery semantics | Prediction accuracy | Recovery semantics | Prediction accuracy | Recovery semantics | |||||||||
| Method | RMSE | MAE | CDE | RTE | RMSE | MAE | CDE | RTE | RMSE | MAE | CDE | RTE | |||
| ARIMAX | 0.099 | 0.078 | 0.423 | 0.356 | 1.355 | 0.112 | 0.086 | 0.307 | 0.478 | 1.342 | 0.128 | 0.098 | 0.229 | 0.579 | 1.557 |
| XGBoost | 0.067 | 0.051 | 0.737 | 0.210 | 1.039 | 0.089 | 0.071 | 0.548 | 0.378 | 1.175 | 0.103 | 0.082 | 0.479 | 0.461 | 1.370 |
| LightGBM | 0.064 | 0.048 | 0.761 | 0.197 | 0.989 | 0.087 | 0.069 | 0.567 | 0.366 | 1.145 | 0.101 | 0.080 | 0.495 | 0.448 | 1.344 |
| DLinear | 0.073 | 0.051 | 0.704 | 0.216 | 1.052 | 0.088 | 0.069 | 0.555 | 0.375 | 1.127 | 0.102 | 0.079 | 0.501 | 0.442 | 1.315 |
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
| Literature | Target | T/ST | Method | Features | Spatial unit | Temporal setting | Interp. | Cross |
| Statistical / probabilistic methods | ||||||||
| Song et al. (2013) | Evacuation mobility | ST | Markov decision process | Historical mobility, home locations, evacuation-network attributes, travel attributes | 10 km grid | 12–81 d | ✓ | ✗ |
| Shimosaka et al. (2015) | Active population count | T | Poisson regression | Historical mobility, temporal covariates, weather covariates, location attributes | 900 m grid | 30–180 d | ✓ | ✗ |
| Song et al. (2014) | Emergency behavior (Individual) | ST | HMM | Historical mobility, important locations, disaster intensity, distance to hazard, government/news reports, population-flow attributes | Regional | 4 h–1 d | ✓ | ✗ |
| Kieu et al. (2025) | POI visits | T | Hawkes process | Historical POI visits, POI categories, dwell time, temporal covariates | Regional | 1–3 h | ✓ | ✗ |
| Machine-learning methods | ||||||||
| Event | Affected tracts | Event start date ( ) | Data collection time frame | Affected users | Disaster intensity | Data sources and roles |
| Hurricanes | ||||||
| Dorian (2019) | 24 | 2019/9/6 | 2019/8/14–2019/9/25 | 60,341 | Rainfall | Hurricane reports: NHC ( NOAA National Hurricane Center, nd ) Hurricane tracks: HURDAT2 ( NOAA National Hurricane Center, 2026 ) Official affected areas: FEMA ( Federal Emergency Management Agency, nd ) Past weather: NOAA PSL ( NOAA Physical Sciences Laboratory, nd ) Forecast weather: GFS ( NOAA National Centers for Environmental Information, nda ) |
| Laura (2020) | 464 | 2020/8/27 | 2020/8/4–2020/9/15 | 510,594 | ||
| Ida (2021) | 403 | 2021/8/29 | 2021/8/6–2021/9/17 | 319,843 | ||
| Nicholas (2021) | 127 | 2021/9/14 | 2021/8/22–2021/10/3 | 218,065 | ||
| Ian (2022) | 341 | 2022/9/28 | 2022/9/5–2022/10/17 | 376,873 | ||
| Category | Type | Metric | Formula/Source | Definition |
| Impact | Magnitude | DI | Maximum post-event mobility loss within the early observation window. | |
| Duration | Timp | Normalized time required to reach the maximum observed impact. | ||
| Integral | ImpPersist | Normalized cumulative impact before the observed minimum; larger values indicate a more persistent impact phase. | ||
| Rate | RelFR | Relative speed at which the observed mobility loss is realized. | ||
| Censoring | Censored | Indicates that the observed minimum occurs at the observation boundary. | ||
| Recovery | Rate | RR | Mean recovery rate from the observed minimum to the observation endpoint. |
Spatiotemporal Forecasting
Spatiotemporal forecasting predicts how quantities change across locations and time, from traffic and weather to disease and environmental systems.
See all 454 papers