ReMoE: An Interpretable Mixture-of-Experts Framework for Human Mobility Recovery Forecasting Across Disasters

Yilin Lyu, Kang Liu, Wei Wang, Qifan Zhang, Ling Yin, Yushi BaiarXiv 2026 · arXiv

TL;DR

ReMoE is a novel "Mixture-of-Experts" framework designed to forecast how human mobility recovers after diverse disasters like hurricanes, floods, and wildfires. By combining specialized, reusable recovery patterns with real-time hazard forecasts, it achieves superior accuracy on unseen disaster types while remaining fully interpretable for emergency responders.

What Problem Does This Paper Address?

When a disaster strikes, human movement patterns—commuting, shopping, socializing—drop sharply and then recover along complex trajectories. Predicting this recovery is vital for allocating resources and assessing community resilience.

However, existing AI models for this task face two major hurdles:

  1. Poor Generalization: Most models are "black boxes" trained on specific events (e.g., a single hurricane). They often fail when applied to a different type of disaster or a new city because they rely on event-specific correlations rather than underlying recovery dynamics.
  2. Lack of Interpretability: Emergency managers need to know why a model predicts a slow recovery for a specific neighborhood. Standard deep learning models offer little transparency, making them difficult to trust in high-stakes decision-making.
Figure 1: Motivation of ReMoE. (a) Event-specific black-box models generalize poorly to unseen disasters. (b) Prior stud- ies reveal recurring and compositional recovery dynamics. (c) ReMoE learns and combines reusable dynamics for inter- pretable cross-disaster forecasting.

Methods

The authors propose ReMoE (Interpretable Mixture-of-Experts), which breaks the forecasting problem into two main modules:

  • Interpretable Recovery-Type Composer (IRTC): Instead of one giant neural network, IRTC uses a set of specialized "experts." Each expert learns a distinct, reusable recovery pattern (e.g., a "rapid rebound" vs. a "long-term deficit"). A feature-guided sparse router then looks at regional context (like social vulnerability and population density) and early mobility data to decide which experts to "listen to" for a given area.
  • Spatiotemporal Pressure-Informed Residual Correction (STP-RC): Since recovery is also driven by ongoing weather or hazard conditions, this module uses future hazard forecasts to "nudge" the base prediction from the experts, accounting for evolving external pressures.

To ensure the experts actually learn different things, the authors use a "trend-residual" design that separates long-term recovery shapes from short-term noise, along with a diversity loss function that penalizes experts for becoming too similar.

Figure 2: Overview of the ReMoE framework.

Datasets

The study utilizes a massive, multi-disaster mobility dataset curated from over 219 million anonymized visit records (via Veraset). It covers 18 major U.S. disaster events between 2019 and 2025, including:

  • 8 Hurricanes (e.g., Ida, Ian, Helene, Milton)
  • 4 Winter Storms (e.g., Uri)
  • 2 Floods (e.g., Vermont 2023)
  • 4 Wildfires (e.g., Lahaina, August Complex)

The data is aggregated at the U.S. Census tract level, covering 5,031 affected spatial units.

Key Insights and Ideas

  • Recovery is Compositional: A key "aha" moment is that while every disaster is unique, the types of recovery are recurring. A neighborhood's trajectory can be seen as a weighted combination of a few fundamental "archetypes."
  • Features Over Latent Space: By using explicit, engineered features (like "impact depth" or "recovery rate") to guide the router rather than abstract mathematical vectors, the model's decisions become "traceable." You can see exactly which regional characteristics led the model to choose a "slow recovery" expert.
  • Robustness Through Specialization: ReMoE actually performs better relative to baselines as the "distribution shift" increases. In other words, its advantage is most pronounced when it has to predict a disaster type it has never seen before, proving that its learned recovery types are truly transferable.

Key Results

ReMoE was tested against 12 state-of-the-art baselines, including statistical methods (ARIMAX), machine learning (XGBoost), and deep learning (iTransformer, PatchTST).

  • Superior Generalization: In the most challenging "Leave-One-Hazard-Type-Out" (LOHTO) setting—where the model is tested on a disaster type it never saw during training—ReMoE reduced error (RMSE) by 8.8% and improved $R^2$ by 0.057 over the best baseline.
  • Semantic Accuracy: It excelled at "recovery-semantic" metrics, reducing timing errors (RTE) by 15.1%, meaning it is much better at predicting when a community will bounce back, not just by how much.
  • Traceable Logic: Post-hoc analysis confirmed that the experts naturally aligned with four empirical recovery types identified in previous social science research, providing a bridge between deep learning and domain knowledge.
Figure 3: Model interpretability. (a) Four empirical recovery types identified through post-hoc clustering. (b) t-SNE projection of router weights colored by above recovery types and “Other” (purple). (c) Mean expert weights by recovery type, revealing soft specialization. (d) Spearman associations between router features and expert weights. (e) Expert-weight compositions across hazard types. (f) Spatial distributions of dominant experts in representative events.

Limitations and Future Work

The authors acknowledge several limitations:

  • Population Representation: The mobility data covers only a subset of the population and is aggregated at the census tract level, which may obscure hyper-local variations.
  • Event Sample Size: While 18 events is a large dataset for this field, more cases—especially "compound events" where multiple disasters overlap—are needed for even more robust conclusions.
  • Causality: The associations found between features and recovery types are statistical; they do not strictly prove that a specific feature caused a specific recovery pattern.

Future work could involve integrating Large Language Models (LLMs) to incorporate textual news reports or social media sentiment into the forecasting process.

Summary updated

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.
BibTeX
@misc{lyu2026remoe,
  author = {Yilin Lyu and Kang Liu and Wei Wang and Qifan Zhang and Ling Yin and Yushi Bai},
  title = {{ReMoE: An Interpretable Mixture-of-Experts Framework for Human Mobility Recovery Forecasting Across Disasters}},
  year = {2026},
  eprint = {2611.22133},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.22133},
}
Figure