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

arXiv cs.LG Papers

Summary

ReMoE 是一个可解释的混合专家框架,用于跨灾害场景的人类移动恢复预测,通过特征引导路由组合可复用的恢复模式,在未见过的灾害事件上取得最佳性能。

arXiv:2611.22133v1 Announce Type: new 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.
Original Article
View Cached Full Text

Cached at: 11/19/26, 03:36 AM

# ReMoE: An Interpretable Mixture-of-Experts Framework for Human Mobility Recovery Forecasting Across Disasters
Source: [https://arxiv.org/html/2611.22133](https://arxiv.org/html/2611.22133)
\\correspondingauthor

CCS:Information systems Geographic information systemsCCS:Information systems Spatial\-temporal systemsYilin LyuAffiliation:School of AI and Liberal Arts,Beijing Normal\-Hong Kong Baptist University,Zhuhai,ChinaAffiliation:Shenzhen University of Advanced Technology,Shenzhen,Chinaemail:[lyuyilin@bnbu\.edu\.cn](mailto:[email protected])Kang LiuAffiliation:Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences,Shenzhen,ChinaAffiliation:Shenzhen University of Advanced Technology,Shenzhen,Chinaemail:[kang\.liu@siat\.ac\.cn](mailto:[email protected]),Wei WangAffiliation:Department of Mechanical Engineering,City University of Hong Kong,Hong Kong,Chinaemail:[wwang326@cityu\.edu\.hk](mailto:[email protected]),Qifan ZhangAffiliation:Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences,Shenzhen,Chinaemail:[622025330022@smail\.nju\.edu\.cn](mailto:[email protected]),Ling YinAffiliation:Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences,Shenzhen,Chinaemail:[yinling@siat\.ac\.cn](mailto:[email protected])andYushi BaiAffiliation:Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences,Shenzhen,Chinaemail:[3250009371@student\.must\.edu\.mo](mailto:[email protected])

© none

###### 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\. The source code is available at: https://github\.com/lvyl9909/ReMoEKDD27\.

###### Keywords:

human mobility recovery, post\-disaster, mixture\-of\-experts, interpretability, cross\-disaster generalization

## 1\.Introduction

![Refer to caption](https://arxiv.org/html/2611.22133v1/figures/fig_introduction.png)Figure 1\.Motivation of ReMoE\. \(a\) Event\-specific black\-box models generalize poorly to unseen disasters\. \(b\) Prior studies reveal recurring and compositional recovery dynamics\. \(c\) ReMoE learns and combines reusable dynamics for interpretable cross\-disaster forecasting\.Human mobility undergoes a decline after disasters and subsequently exhibits a complex recovery trajectory\. This process provides a timely view of how urban functions are affected and restored, while revealing substantial spatial heterogeneity in recovery across communities\([Yabe et al\., 2020](https://arxiv.org/html/2611.22133#bib.bib72)\)\. However, retrospective observation is only part of the problem\. In rapidly evolving scenarios, proactive decision\-making requires an estimate of how mobility will evolve over the next several days\. Thus, accurately predicting this dynamic recovery process holds paramount importance for informing post\-disaster recovery strategies, including prioritizing resources for communities expected to experience prolonged disruption\([Wang et al\., 2023](https://arxiv.org/html/2611.22133#bib.bib69)\)\.

However, post\-disaster mobility recovery forecasting faces two long\-standing challenges\.First, cross\-disaster generalization remains limited\. Recovery dynamics vary widely in their nature, magnitude, and duration from event to event, in large part because the events themselves differ so greatly\. In addition, the heterogeneity of mobility change can vary at different geographies, manifested by the economy, demography, and physical and social infrastructure\([Hong et al\., 2021a](https://arxiv.org/html/2611.22133#bib.bib19)\)\. Consequently, many existing studies develop ad hoc models for a single disaster type or study area\([Wang et al\., 2022](https://arxiv.org/html/2611.22133#bib.bib71);[Tang et al\., 2024](https://arxiv.org/html/2611.22133#bib.bib65);[Jiang et al\., 2023a](https://arxiv.org/html/2611.22133#bib.bib26)\)\. Although informative in their settings, these models often rely on event\- or city\-specific correlations that may not generalize to unseen disasters or cities, limiting rapid deployment without enough local calibration data\.Second, model interpretability remains limited\. Although a growing body of research has achieved high performance, less attention has been paid to explaining model predictions and grounding them in underlying recovery mechanisms\([Li et al\., 2024](https://arxiv.org/html/2611.22133#bib.bib33);[Yang et al\., 2025a](https://arxiv.org/html/2611.22133#bib.bib74)\)\. However, such transparency is critical for decision support, as decision\-makers must understand and evaluate a forecast rather than rely on a black\-box output\([Rudin, 2019](https://arxiv.org/html/2611.22133#bib.bib58)\)\.

Recent post hoc studies have shown that mobility recovery can follow diverse but recurring patterns across disasters\([Tang et al\., 2023](https://arxiv.org/html/2611.22133#bib.bib66);[Hsu and Mostafavi, 2025](https://arxiv.org/html/2611.22133#bib.bib22);[Li et al\., 2026](https://arxiv.org/html/2611.22133#bib.bib34)\)\. For example,[Hsu and Mostafavi \(2025\)](https://arxiv.org/html/2611.22133#bib.bib22)identified three archetypes from more than 2,000 empirical resilience curves, suggesting that different disasters may share three reusable recovery types\. Beyond such discrete types, the actual regional recovery trajectory aggregates heterogeneous local processes and may therefore be better represented as flexible compositions of multiple latent recovery types\([Tang et al\., 2023](https://arxiv.org/html/2611.22133#bib.bib66);[Jiang et al\., 2023b](https://arxiv.org/html/2611.22133#bib.bib27)\)\. At the same time, mixture\-of\-experts \(MoE\) models have demonstrated strong performance in learning heterogeneous data representations\. They employ specialized “experts” to capture heterogeneous dynamics and adaptively combine their outputs through a routing mechanism\([Shazeer et al\., 2017](https://arxiv.org/html/2611.22133#bib.bib59)\)\. This compositional paradigm may be naturally aligned with regional recovery heterogeneity; e\.g\., specialized experts can capture reusable recovery types, while router weights combine them to represent region\-specific recovery processes\. This correspondence may consequently enable recovery knowledge to transfer to unseen disasters\.

Motivated by these findings, we propose to leverage MoE architecture as a natural fit for predicting heterogeneous recovery dynamics\. Yet, the black\-box nature of standard MoE models obscures how input shapes expert specialization, making semantically grounded generalization difficult\([Karniadakis et al\., 2021](https://arxiv.org/html/2611.22133#bib.bib28);[Huang et al\., 2026](https://arxiv.org/html/2611.22133#bib.bib23)\)\. To address this limitation, we proposeReMoE, an interpretable mixture\-of\-experts framework for forecasting human mobility recovery across disasters\. ReMoE consists of two complementary modules\. First, the Interpretable Recovery\-Type Composer \(IRTC\) provides the main source of interpretability and cross\-disaster transfer\. It uses engineered features derived from early mobility observations and regional context to guide routing, making expert allocation traceable and selective\. Also, its trend\-residual expert design separates underlying recovery patterns from local fluctuations to promote stable expert specialization\. Second, the Spatiotemporal Pressure\-Informed Residual Correction \(STP\-RC\) module incorporates forecast hazard conditions as exogenous signals to improve model performance\. The main contributions can be summarized as follows:

- •To the best of our knowledge, we are the first to forecast transferable human mobility recovery across unseen events and hazard types\. The idea of leveraging MoE to learn and compose heterogeneous yet reusable recovery dynamics supports this generalization\.
- •We develop an interpretable MoE framework that integrates a feature\-guided sparse router with trend\-residual experts, enabling post\-hoc explanations of routing decisions and intrinsic interpretation of expert specialization\.
- •We curate a large\-scale mobility dataset covering 18 disasters, over 219 million visit records, and 5,031 affected spatial units\. Extensive experiments demonstrate ReMoE’s superior accuracy, interpretability, and robustness, particularly under unseen\-event and unseen\-hazard\-type settings\.

## 2\.Preliminaries

Definition 2\.1 \(Normalized Region\-Wise Mobility Sequence\)\.For disaster eventee, letℛe=\{r1,…,rN\}\\mathcal\{R\}\_\{e\}=\\\{r\_\{1\},\\ldots,r\_\{N\}\\\}denote itsNNaffected regions,𝒯e=\{τ1,…,τT\}\\mathcal\{T\}\_\{e\}=\\\{\\tau\_\{1\},\\ldots,\\tau\_\{T\}\\\}denoteTTconsecutive daily intervals, andτ0e\\tau\_\{0\}^\{e\}denote the event date\. Letfi​j,e​\(τ\)f\_\{ij,e\}\(\\tau\)be the number of directed movements from regionrir\_\{i\}to regionrjr\_\{j\}on dayτ\\tau\. The total inter\-region mobility ofrir\_\{i\}and its normalized value are

\(1\)mi,e​\(τ\)=∑j≠i\[fj​i,e​\(τ\)\+fi​j,e​\(τ\)\],m~i,e​\(τ\)=mi,e​\(τ\)m¯i,epre\.m\_\{i,e\}\(\\tau\)=\\sum\_\{j\\neq i\}\\left\[f\_\{ji,e\}\(\\tau\)\+f\_\{ij,e\}\(\\tau\)\\right\],\\qquad\\widetilde\{m\}\_\{i,e\}\(\\tau\)=\\frac\{m\_\{i,e\}\(\\tau\)\}\{\\overline\{m\}^\{\\mathrm\{pre\}\}\_\{i,e\}\}\.Here,m¯i,epre\\overline\{m\}^\{\\mathrm\{pre\}\}\_\{i,e\}denotes the pre\-event baseline, such thatm~i,e​\(τ\)=1\\widetilde\{m\}\_\{i,e\}\(\\tau\)=1represents normal mobility\. The corresponding normalized sequence over𝒯e\\mathcal\{T\}\_\{e\}is𝐦~i,e=\[m~i,e​\(τ1\),…,m~i,e​\(τT\)\]∈ℝT\\widetilde\{\\mathbf\{m\}\}\_\{i,e\}=\[\\widetilde\{m\}\_\{i,e\}\(\\tau\_\{1\}\),\\ldots,\\widetilde\{m\}\_\{i,e\}\(\\tau\_\{T\}\)\]\\in\\mathbb\{R\}^\{T\}\.

In this study, each region corresponds to a U\.S\. Census tract; intra\-region visitsfi​i,e​\(τ\)f\_\{ii,e\}\(\\tau\)are excluded, and origin–destination interactions are not explicitly modeled\. Moreover, the normalized mobility is processed using a three\-day causal moving average:

\(2\)m~i,esm​\(τ\)=m~i,e​\(τ\)\+m~i,e​\(τ−1\)\+m~i,e​\(τ−2\)3\.\\widetilde\{m\}^\{\\mathrm\{sm\}\}\_\{i,e\}\(\\tau\)=\\frac\{\\widetilde\{m\}\_\{i,e\}\(\\tau\)\+\\widetilde\{m\}\_\{i,e\}\(\\tau\-1\)\+\\widetilde\{m\}\_\{i,e\}\(\\tau\-2\)\}\{3\}\.Our model input is derived from this smoothed sequence by restricting it to a specified observation window\. LetTpreT\_\{\\mathrm\{pre\}\}andTpostT\_\{\\mathrm\{post\}\}denote the numbers of observed days before and afterτ0e\\tau\_\{0\}^\{e\}, respectively\. The resulting input sequence is

\(3\)𝐦~i,ein=\[m~i,esm​\(τ0e−Tpre\),…,m~i,esm​\(τ0e\+Tpost\)\]\.\\widetilde\{\\mathbf\{m\}\}^\{\\mathrm\{in\}\}\_\{i,e\}=\\left\[\\widetilde\{m\}^\{\\mathrm\{sm\}\}\_\{i,e\}\(\\tau\_\{0\}^\{e\}\-T\_\{\\mathrm\{pre\}\}\),\\ldots,\\widetilde\{m\}^\{\\mathrm\{sm\}\}\_\{i,e\}\(\\tau\_\{0\}^\{e\}\+T\_\{\\mathrm\{post\}\}\)\\right\]\.
Definition 2\.2 \(Spatiotemporal Pressure\)\.Following the Spatiotemporal \(ST\) Decay Model proposed by Li et al\.\([Li et al\., 2022](https://arxiv.org/html/2611.22133#bib.bib35)\), we characterize the external hazard pressure in the regionrir\_\{i\}as

\(4\)Pi,e​\(τ\)=ke​\(τ\)​∑j∈𝒩⁡\(i\)∪\{i\}wi​j​Hj,e​\(τ\),P\_\{i,e\}\(\\tau\)=k\_\{e\}\(\\tau\)\\sum\_\{j\\in\\mathcal\{N\}\(i\)\\cup\\\{i\\\}\}w\_\{ij\}H\_\{j,e\}\(\\tau\),whereHj,e​\(τ\)H\_\{j,e\}\(\\tau\)denotes the hazard intensity of regionrjr\_\{j\},wi​jw\_\{ij\}captures neighborhood influence, andke​\(τ\)=k0​e−α⁡\(τ−τ0e\)k\_\{e\}\(\\tau\)=k\_\{0\}e^\{\-\\alpha\(\\tau\-\\tau\_\{0\}^\{e\}\)\}captures temporal decay after the event\. We map this quantity to a bounded spatiotemporal pressure \(STP\) score:

\(5\)STPi,e⁡\(τ\)=Pi,e​\(τ\)1\+Pi,e​\(τ\)∈\[0,1\)\.\\operatorname\{STP\}\_\{i,e\}\(\\tau\)=\\frac\{P\_\{i,e\}\(\\tau\)\}\{1\+P\_\{i,e\}\(\\tau\)\}\\in\[0,1\)\.A larger STP indicates a stronger external hazard pressure\. In this study, observed STP is calculated from hazard conditions within the observation window, whereas the future STP sequence𝐒𝐓𝐏i,efuture\\mathbf\{STP\}^\{\\mathrm\{future\}\}\_\{i,e\}is derived from hazard forecasts available at the forecast origin\.

Problem Formulation\.For each affected regionrir\_\{i\}under disaster eventee, given its observed normalized mobility sequence𝐦~i,ein\\widetilde\{\\mathbf\{m\}\}^\{\\mathrm\{in\}\}\_\{i,e\}, contextual features𝐂i,e\\mathbf\{C\}\_\{i,e\}, and forecast spatiotemporal pressure𝐒𝐓𝐏i,efuture\\mathbf\{STP\}^\{\\mathrm\{future\}\}\_\{i,e\}, we aim to predict its subsequentHH\-day mobility recovery:

\(6\)𝐦^i,efuture=ℱθ​\(𝐦~i,ein,𝐂i,e,𝐒𝐓𝐏i,efuture\)\.\\widehat\{\\mathbf\{m\}\}^\{\\mathrm\{future\}\}\_\{i,e\}=\\mathcal\{F\}\_\{\\theta\}\\left\(\\widetilde\{\\mathbf\{m\}\}^\{\\mathrm\{in\}\}\_\{i,e\},\\mathbf\{C\}\_\{i,e\},\\mathbf\{STP\}^\{\\mathrm\{future\}\}\_\{i,e\}\\right\)\.The forecasting functionℱθ\\mathcal\{F\}\_\{\\theta\}is learned jointly across multiple events and hazard types to support generalization to unseen disasters\.

![Diagram of the ReMoE framework. Mobility-derived recovery features and hazard-pressure and regional features enter a top-2 router, which combines forecasts from trend-residual experts into a Base trajectory. A BiLSTM then uses the Base trajectory and Future ST Pressure to produce a residual correction and the final mobility recovery forecast.](https://arxiv.org/html/2611.22133v1/figures/fig_method_overview.png)Figure 2\.Overview of the ReMoE framework\.Diagram of the ReMoE framework\. Mobility\-derived recovery features and hazard\-pressure and regional features enter a top\-2 router, which combines forecasts from trend\-residual experts into a Base trajectory\. A BiLSTM then uses the Base trajectory and Future ST Pressure to produce a residual correction and the final mobility recovery forecast\.
## 3\.Methodology

### 3\.1\.Overview

The architecture of the ReMoE framework is shown in Figure[2](https://arxiv.org/html/2611.22133#acmlabel1)\. It comprises two complementary modules that integrate process\-informed knowledge with data\-driven learning: \(1\) the Interpretable Recovery\-Type Composer \(IRTC\) and \(2\) the Spatiotemporal Pressure\-Informed Residual Correction \(STP\-RC\)\. IRTC implements an MoE architecture with a feature\-guided sparse router and a semantics\-oriented expert design to learn a base recovery pattern while encouraging experts to develop distinct and stable semantic roles, thereby enhancing model interpretability\. STP\-RC then uses the forecast hazard conditions to estimate residual corrections to this base trajectory and produce the final forecast\. Together, the two modules provide distinct yet complementary capabilities\.

### 3\.2\.Interpretable Recovery\-Type Composer

#### 3\.2\.1\.Feature\-Guided Sparse Router\.

A growing body of work suggests that MoE performance depends not only on the capacity of experts, but also on how effectively the router induces meaningful specialization among them\([Guo et al\., 2026](https://arxiv.org/html/2611.22133#bib.bib15)\)\. Unlike conventional MoE routers operating on latent representations, our router uses explicit features which have been shown to be informative for distinguishing subsequent recovery types\([Yabe et al\., 2020](https://arxiv.org/html/2611.22133#bib.bib72);[Rachunok and Nateghi, 2021](https://arxiv.org/html/2611.22133#bib.bib54)\), encouraging clearer expert differentiation and making routing decisions traceable\([Swamy et al\., 2025](https://arxiv.org/html/2611.22133#bib.bib64)\)\. We further adopt sparse Top\-KKrouting strategy, which assigns each sample to only a small subset of experts and reduces excessive functional overlap among them\.

From the observed normalized mobility sequence and contextual conditions, we construct 16 features in two categories: recovery\-process features and contextual features\. Following prior resilience\-curve studies\([Hillebrand et al\., 2018](https://arxiv.org/html/2611.22133#bib.bib17);[Poulin and Kane, 2021](https://arxiv.org/html/2611.22133#bib.bib52)\), these recovery\-process features characterize properties such as magnitude, duration, integral, rate, and stability\. We derive them separately from the impact and early\-recovery periods because these stages provide complementary signals for subsequent recovery\([Yabe et al\., 2020](https://arxiv.org/html/2611.22133#bib.bib72)\)\. Contextual features distinguish similar mobility patterns using observed STP and regional conditions \(e\.g\., social vulnerability, population density, built\-up share, and network centrality\)\. After redundancy analysis, the final set contains five impact, six recovery, and five contextual features:

\(7\)𝐅i,e=\[Φimp​\(𝐦~i,ein\),Φrec​\(𝐦~i,ein\),𝐂i,e\]∈ℝ16,\\mathbf\{F\}\_\{i,e\}=\\left\[\\Phi^\{\\mathrm\{imp\}\}\\\!\\left\(\\widetilde\{\\mathbf\{m\}\}^\{\\mathrm\{in\}\}\_\{i,e\}\\right\),\\Phi^\{\\mathrm\{rec\}\}\\\!\\left\(\\widetilde\{\\mathbf\{m\}\}^\{\\mathrm\{in\}\}\_\{i,e\}\\right\),\\mathbf\{C\}\_\{i,e\}\\right\]\\in\\mathbb\{R\}^\{16\},whereΦimp​\(⋅\)∈ℝ5\\Phi^\{\\mathrm\{imp\}\}\(\\cdot\)\\in\\mathbb\{R\}^\{5\},Φrec​\(⋅\)∈ℝ6\\Phi^\{\\mathrm\{rec\}\}\(\\cdot\)\\in\\mathbb\{R\}^\{6\}, and𝐂i,e∈ℝ5\\mathbf\{C\}\_\{i,e\}\\in\\mathbb\{R\}^\{5\}denote the impact, recovery, and contextual feature categories, respectively\. Definitions and filtering details are provided in the Appendix[C](https://arxiv.org/html/2611.22133#A3)\.

The expert weightsgi,e,jg\_\{i,e,j\}are obtained through a softmax gating mechanism followed by top\-KKselection and renormalization:

\(8\)gi,e,j=\{g~i,e,j∑ℓ∈𝒦i,eg~i,e,ℓ,j∈𝒦i,e,0,otherwise,g\_\{i,e,j\}=\\begin\{cases\}\\displaystyle\\frac\{\\widetilde\{g\}\_\{i,e,j\}\}\{\\sum\_\{\\ell\\in\\mathcal\{K\}\_\{i,e\}\}\\widetilde\{g\}\_\{i,e,\\ell\}\},&j\\in\\mathcal\{K\}\_\{i,e\},\\\\\[6\.0pt\] 0,&\\text\{otherwise\},\\end\{cases\}whereg~i,e,j\\widetilde\{g\}\_\{i,e,j\}is the dense gating probability for expertjj, and𝒦i,e\\mathcal\{K\}\_\{i,e\}contains the indices of its top\-KKprobabilities\. Here, we setK=2K=2\.

#### 3\.2\.2\.Trend\-Residual Experts\.

Post\-disaster mobility recovery is heterogeneous, with affected sub\-areas exhibiting distinct recovery processes\([Hsu and Mostafavi, 2025](https://arxiv.org/html/2611.22133#bib.bib22);[Tang et al\., 2023](https://arxiv.org/html/2611.22133#bib.bib66)\)\. A multi\-expert structure is therefore employed to capture these diverse dynamics\. Ideally, each expert should specialize in a single recovery regime, allowing the weighted expert composition to provide semantic interpretability\([Yang et al\., 2025b](https://arxiv.org/html/2611.22133#bib.bib75)\)\. However, when experts operate directly on raw mobility sequences, long\-term recovery shapes and short\-term fluctuations are entangled\. Consequently, different experts may learn different mixtures of these components, obscuring their functional roles\.

To promote clearer specialization, we explicitly decompose the input to each expert into a moving\-average trend and a remainder:

\(9\)𝐓𝐫𝐞𝐧𝐝i,e=MAL⁡\(𝐦~i,ein\),𝐑𝐞𝐬𝐢𝐝𝐮𝐚𝐥i,e=𝐦~i,ein−𝐓𝐫𝐞𝐧𝐝i,e,\\mathbf\{Trend\}\_\{i,e\}=\\operatorname\{MA\}\_\{L\}\\left\(\\widetilde\{\\mathbf\{m\}\}^\{\\mathrm\{in\}\}\_\{i,e\}\\right\),\\qquad\\mathbf\{Residual\}\_\{i,e\}=\\widetilde\{\\mathbf\{m\}\}^\{\\mathrm\{in\}\}\_\{i,e\}\-\\mathbf\{Trend\}\_\{i,e\},whereL=3L=3is the moving\-average window\. Expertkkthen produces anHH\-day trajectory:

\(10\)𝐄𝐱𝐩𝐞𝐫𝐭i,e\(k\)=ℰktrend​\(𝐓𝐫𝐞𝐧𝐝i,e\)\+σ⁡\(ak\)​ℰkres​\(𝐑𝐞𝐬𝐢𝐝𝐮𝐚𝐥i,e\)∈ℝH,\\mathbf\{Expert\}^\{\(k\)\}\_\{i,e\}=\\mathcal\{E\}\_\{k\}^\{\\mathrm\{trend\}\}\\left\(\\mathbf\{Trend\}\_\{i,e\}\\right\)\+\\sigma\(a\_\{k\}\)\\mathcal\{E\}\_\{k\}^\{\\mathrm\{res\}\}\\left\(\\mathbf\{Residual\}\_\{i,e\}\\right\)\\in\\mathbb\{R\}^\{H\},where the two components are MLPs,σ⁡\(ak\)\\sigma\(a\_\{k\}\)controls the residual contribution\. The final IRTC output for regioniiis a weighted sum:

\(11\)𝐁𝐚𝐬𝐞i,e=∑k=1Egi,e,k​𝐄𝐱𝐩𝐞𝐫𝐭i,e\(k\)\.\\mathbf\{Base\}\_\{i,e\}=\\sum\_\{k=1\}^\{E\}g\_\{i,e,k\}\\mathbf\{Expert\}^\{\(k\)\}\_\{i,e\}\.
Note that𝐁𝐚𝐬𝐞i,e∈ℝH\\mathbf\{Base\}\_\{i,e\}\\in\\mathbb\{R\}^\{H\}is a learned representation of basic recovery pattern whose length matches the prediction horizon\. It is subsequently forwarded to STP\-RC for correction\.

### 3\.3\.Spatiotemporal Pressure\-Informed Residual Correction

IRTC captures a basic recovery pattern from the observed sequences, but foreseeable future hazard conditions may cause additional deviations\([Lim et al\., 2021](https://arxiv.org/html/2611.22133#bib.bib36)\)\. To include these influences, we employ a recurrent residual corrector to translate the evolving forecast STP into adjustments to theBase\. Let𝐒𝐓𝐏i,efuture=\[S​T​Pi,efuture​\(1\),…,S​T​Pi,efuture​\(H\)\]⊤\\mathbf\{STP\}^\{\\mathrm\{future\}\}\_\{i,e\}=\[STP^\{\\mathrm\{future\}\}\_\{i,e\}\(1\),\\ldots,STP^\{\\mathrm\{future\}\}\_\{i,e\}\(H\)\]^\{\\top\}be the future STP sequence\. It is worth noting that𝐒𝐓𝐏i,efuture\\mathbf\{STP\}^\{\\mathrm\{future\}\}\_\{i,e\}is constructed from hazard forecasts issued no later than the forecast origin and converted into tract\-level STP values\. It is concatenated with the𝐁𝐚𝐬𝐞i,e\\mathbf\{Base\}\_\{i,e\}as

\(12\)𝐗i,e=Concat⁡\[sg⁡\(𝐁𝐚𝐬𝐞i,e\),𝐒𝐓𝐏i,efuture\]∈ℝH×2,\\mathbf\{X\}\_\{i,e\}=\\operatorname\{Concat\}\\left\[\\operatorname\{sg\}\\\!\\left\(\\mathbf\{Base\}\_\{i,e\}\\right\),\\mathbf\{STP\}^\{\\mathrm\{future\}\}\_\{i,e\}\\right\]\\in\\mathbb\{R\}^\{H\\times 2\},wheresg⁡\(⋅\)\\operatorname\{sg\}\(\\cdot\)denotes the stop\-gradient operation, allowing STP\-RC to use theBasewithout propagating correction gradients back to the experts, therefore IRTC remains trained because its output directly contributes to the final prediction\([Chen and He, 2021](https://arxiv.org/html/2611.22133#bib.bib10)\)\. A bidirectional LSTM then processes𝐗i,e\\mathbf\{X\}\_\{i,e\}and estimates the pressure\-driven residual:

\(13\)𝚫i,e=σ⁡\(af\)​\[BiLSTM⁡\(𝐗i,e\)​𝐰f\+bf\],\\boldsymbol\{\\Delta\}\_\{i,e\}=\\sigma\(a\_\{f\}\)\\left\[\\operatorname\{BiLSTM\}\(\\mathbf\{X\}\_\{i,e\}\)\\mathbf\{w\}\_\{f\}\+b\_\{f\}\\right\],where𝐰f∈ℝ2​dh\\mathbf\{w\}\_\{f\}\\in\\mathbb\{R\}^\{2d\_\{h\}\}andbf∈ℝb\_\{f\}\\in\\mathbb\{R\}define a shared linear projection that maps each bidirectional hidden state to a scalar correction\.σ⁡\(af\)\\sigma\(a\_\{f\}\)controls the overall correction magnitude\. The final prediction is

\(14\)𝐦^i,efuture=𝐁𝐚𝐬𝐞i,e\+𝚫i,e\.\\widehat\{\\mathbf\{m\}\}^\{\\mathrm\{future\}\}\_\{i,e\}=\\mathbf\{Base\}\_\{i,e\}\+\\boldsymbol\{\\Delta\}\_\{i,e\}\.
STP\-RC sequentially adjusts theBaseto evolving forecast hazard conditions, improving predictive accuracy\.

### 3\.4\.Model Training

ReMoE is trained end\-to\-end by minimizing a loss function that balances prediction accuracy, router utilization and expert diversity\. First, letℬ⊂𝒟train\\mathcal\{B\}\\subset\\mathcal\{D\}\_\{\\mathrm\{train\}\}denote a mini\-batch of tract–event samples\. For each sample\(i,e\)∈ℬ\(i,e\)\\in\\mathcal\{B\}, let𝐦~i,efuture∈ℝH\\widetilde\{\\mathbf\{m\}\}^\{\\mathrm\{future\}\}\_\{i,e\}\\in\\mathbb\{R\}^\{H\}denote the ground\-truth normalized mobility trajectory and𝐦^i,efuture∈ℝH\\widehat\{\\mathbf\{m\}\}^\{\\mathrm\{future\}\}\_\{i,e\}\\in\\mathbb\{R\}^\{H\}denote the corresponding ReMoE output\. The trajectory fidelity loss is

\(15\)ℒpred=1\|ℬ\|​H​∑\(i,e\)∈ℬ‖𝐦^i,efuture−𝐦~i,efuture‖22\.\\mathcal\{L\}\_\{\\mathrm\{pred\}\}=\\frac\{1\}\{\|\\mathcal\{B\}\|H\}\\sum\_\{\(i,e\)\\in\\mathcal\{B\}\}\\left\\\|\\widehat\{\\mathbf\{m\}\}^\{\\mathrm\{future\}\}\_\{i,e\}\-\\widetilde\{\\mathbf\{m\}\}^\{\\mathrm\{future\}\}\_\{i,e\}\\right\\\|\_\{2\}^\{2\}\.Because𝐦^i,efuture=𝐁𝐚𝐬𝐞i,e\+𝚫i,e\\widehat\{\\mathbf\{m\}\}^\{\\mathrm\{future\}\}\_\{i,e\}=\\mathbf\{Base\}\_\{i,e\}\+\\boldsymbol\{\\Delta\}\_\{i,e\}, this objective jointly trains IRTC and STP\-RC without requiring a separate correction loss\.

Second, MoE may suffer from “expert collapse”, where one expert is repeatedly selected\. We therefore calculate the mean router weight assigned to expertkkwithin each mini\-batch:

\(16\)Ak=1\|ℬ\|∑\(i,e\)∈ℬgi,e,k,k=1,…,E\.A\_\{k\}=\\frac\{1\}\{\|\\mathcal\{B\}\|\}\\sum\_\{\(i,e\)\\in\\mathcal\{B\}\}g\_\{i,e,k\},\\qquad k=1,\\ldots,E\.Here,gi,e,kg\_\{i,e,k\}andEEis the expert weight and the number of experts, respectively\. Balanced usage is encouraged by penalizing the coefficient of variation across experts:

\(17\)ℒrouter=Var⁡\(A1,…,AE\)Mean⁡\(A1,…,AE\)2\+ϵ,\\mathcal\{L\}\_\{\\mathrm\{router\}\}=\\frac\{\\operatorname\{Var\}\(A\_\{1\},\\ldots,A\_\{E\}\)\}\{\\operatorname\{Mean\}\(A\_\{1\},\\ldots,A\_\{E\}\)^\{2\}\+\\epsilon\},whereϵ\\epsilonis a positive constant for numerical stability\.

Third, to discourage experts from learning similar recovery dynamics, we calculate their batch\-mean candidate trajectories:

\(18\)𝐳¯k=1\|ℬ\|​∑\(i,e\)∈ℬ𝐄𝐱𝐩𝐞𝐫𝐭i,e\(k\)\.\\overline\{\\mathbf\{z\}\}\_\{k\}=\\frac\{1\}\{\|\\mathcal\{B\}\|\}\\sum\_\{\(i,e\)\\in\\mathcal\{B\}\}\\mathbf\{Expert\}^\{\(k\)\}\_\{i,e\}\.
We then compare the mean\-centered shapes of their batch\-averaged trajectories and penalize pairwise cosine similarities above margin:

\(19\)ℒdiv=2E⁡\(E−1\)​∑1≤p<q≤E\[max⁡\(0,simcos⁡\(ctr⁡\(𝐳¯p\),ctr⁡\(𝐳¯q\)\)−ρ\)\]2,\\mathcal\{L\}\_\{\\mathrm\{div\}\}=\\frac\{2\}\{E\(E\-1\)\}\\sum\_\{1\\leq p<q\\leq E\}\\left\[\\max\\left\(0,\\operatorname\{sim\}\_\{\\cos\}\\left\(\\operatorname\{ctr\}\(\\overline\{\\mathbf\{z\}\}\_\{p\}\),\\operatorname\{ctr\}\(\\overline\{\\mathbf\{z\}\}\_\{q\}\)\\right\)\-\\rho\\right\)\\right\]^\{2\},wherectr⁡\(⋅\)\\operatorname\{ctr\}\(\\cdot\)removes the temporal mean,simcos⁡\(⋅,⋅\)\\operatorname\{sim\}\_\{\\cos\}\(\\cdot,\\cdot\)denotes cosine similarity, andρ\\rhois the maximum similarity allowed without penalty\.

The total training objective is

\(20\)ℒtotal=ℒpred\+λr​ℒrouter\+λd​ℒdiv,\\mathcal\{L\}\_\{\\mathrm\{total\}\}=\\mathcal\{L\}\_\{\\mathrm\{pred\}\}\+\\lambda\_\{\\mathrm\{r\}\}\\mathcal\{L\}\_\{\\mathrm\{router\}\}\+\\lambda\_\{\\mathrm\{d\}\}\\mathcal\{L\}\_\{\\mathrm\{div\}\},whereλr\\lambda\_\{\\mathrm\{r\}\}andλd\\lambda\_\{\\mathrm\{d\}\}weight router balancing and expert diversity\.

## 4\.Experiments

We evaluate ReMoE through the following evaluation questions:

- •EQ1:How accurately does ReMoE forecast post\-disaster mobility recovery, and how well does it generalize to unseen events and hazard types?→\\rightarrowSec\.[4\.2](https://arxiv.org/html/2611.22133#S4.SS2)
- •EQ2:How does ReMoE provide interpretable insights into its learned experts and routing behavior?→\\rightarrowSec\.[4\.3](https://arxiv.org/html/2611.22133#S4.SS3)
- •EQ3:How robust is ReMoE under uncertainty and different model settings?→\\rightarrowSec\.[4\.4](https://arxiv.org/html/2611.22133#S4.SS4)

### 4\.1\.Experimental Setup

#### 4\.1\.1\.Datasets\.

To support cross\-disaster forecasting and evaluate generalization to unseen hazards, we construct a multi\-disaster mobility dataset from over 219 million visit records provided by Veraset, a company that aggregates anonymized location data from multiple apps on Apple and Android platforms\. The dataset contains 18 events between 2019 and 2025, including eight hurricanes, two floods, four winter storms, and four wildfires, covering 5031 affected census tracts\. Appendix[B\.1](https://arxiv.org/html/2611.22133#A2.SS1)shows the details of datasets\.

Using this dataset, we evaluate ReMoE under three settings with progressively increasing distribution shift\. Thein\-distribution \(ID\)setting randomly partitions tract–event samples while retaining all events and hazard types across the data splits\. The other two settings evaluateout\-of\-distribution \(OOD\)generalization\. First, to evaluate generalization to unseen events, we construct eight stratifiedleave\-one\-event\-out \(LOEO\)folds, with one event from each hazard type withheld for testing in each fold and the remaining 14 events used for training\. For hazard types with fewer than eight events, test events are reused across folds following a balanced rotation\. Second, to test transferability to an unseen hazard type, we further introduce a four\-foldleave\-one\-hazard\-type\-out \(LOHTO\)setting, where all events of one hazard type are withheld in each fold\. Results are aggregated across folds\.

#### 4\.1\.2\.Experiment Setting\.

We use the event\-aligned mobility sequence fromt0−9t\_\{0\}\-9tot0\+5t\_\{0\}\+5as the observation window to forecast the subsequent seven\-day recovery trajectory fromt0\+6t\_\{0\}\+6tot0\+12t\_\{0\}\+12\(see Appendix[B\.1\.3](https://arxiv.org/html/2611.22133#A2.SS1.SSS3)for details\)\. For ID evaluation, census tracts are partitioned into training, validation, and test sets at a ratio of 7:2:1, whereas LOEO and LOHTO use the fixed event\-level folds described above\. To mitigate event\-size imbalance, each training sample is weighted inversely by the number of tracts in its event\. ReMoE uses a two\-layer MLP router and five experts\. The model is trained end\-to\-end with AdamW using a learning rate of1\.2×10−31\.2\\times 10^\{\-3\}, weight decay of2×10−42\\times 10^\{\-4\}, batch size 128, and random seed 42\. Models are selected according to validation RMSE, and all baselines use the same observation and forecasting horizons\.

#### 4\.1\.3\.Evaluation Metrics\.

We evaluate forecasts with three standard metrics, RMSE, MAE, andR2R^\{2\}\. However, pointwise errors can be similar for forecasts that recover early or late, so we introduce two recovery\-semantics metrics: the cumulative deficit error \(CDE\) measures the cumulative error over the prediction horizon, whereas the recovery timing error \(RTE\) measures the error in when the mobility deficit is concentrated, reported in days\. For a mobility sequence𝐦=\(m1,…,mH\)\\mathbf\{m\}=\(m\_\{1\},\\ldots,m\_\{H\}\), we define the total deficit and its temporal centroid as

\(21\)D⁡\(𝐦\)=∑h=1Hmax⁡\(1−mh,0\),C⁡\(𝐦\)=∑h=1Hh​max⁡\(1−mh,0\)D⁡\(𝐦\)\+ϵ,D\(\\mathbf\{m\}\)=\\sum\_\{h=1\}^\{H\}\\max\(1\-m\_\{h\},0\),\\qquad C\(\\mathbf\{m\}\)=\\frac\{\\sum\_\{h=1\}^\{H\}h\\max\(1\-m\_\{h\},0\)\}\{D\(\\mathbf\{m\}\)\+\\epsilon\},whereϵ\\epsilonis a small positive constant for numerical stability\.

To prevent large events from dominating the evaluation, all metrics are event\-balanced\. Letℰ\\mathcal\{E\}denote the set of evaluated events andNeN\_\{e\}the number of test tracts in eventee\. Each tract–event sample receives the weight

\(22\)wi,e=1\|ℰ\|​Ne,w\_\{i,e\}=\\frac\{1\}\{\|\\mathcal\{E\}\|N\_\{e\}\},so that all events contribute equally\. RMSE, MAE, andR2R^\{2\}are computed with these weights, and

\(23\)CDE=∑e∈ℰ∑i=1Newi,e​\|D⁡\(𝐦^i,efuture\)−D⁡\(𝐦~i,efuture\)\|,\\operatorname\{CDE\}=\\sum\_\{e\\in\\mathcal\{E\}\}\\sum\_\{i=1\}^\{N\_\{e\}\}w\_\{i,e\}\\left\|D\\\!\\left\(\\widehat\{\\mathbf\{m\}\}^\{\\mathrm\{future\}\}\_\{i,e\}\\right\)\-D\\\!\\left\(\\widetilde\{\\mathbf\{m\}\}^\{\\mathrm\{future\}\}\_\{i,e\}\\right\)\\right\|,\(24\)RTE=∑e∈ℰ∑i=1Newi,e​\|C⁡\(𝐦^i,efuture\)−C⁡\(𝐦~i,efuture\)\|\.\\operatorname\{RTE\}=\\sum\_\{e\\in\\mathcal\{E\}\}\\sum\_\{i=1\}^\{N\_\{e\}\}w\_\{i,e\}\\left\|C\\\!\\left\(\\widehat\{\\mathbf\{m\}\}^\{\\mathrm\{future\}\}\_\{i,e\}\\right\)\-C\\\!\\left\(\\widetilde\{\\mathbf\{m\}\}^\{\\mathrm\{future\}\}\_\{i,e\}\\right\)\\right\|\.Lower values are better for all metrics exceptR2R^\{2\}\.

#### 4\.1\.4\.Baselines and variants\.

We compare ReMoE against twelve baselines spanning four families: statistical methods \(ARIMAX\([Box et al\., 2015](https://arxiv.org/html/2611.22133#bib.bib5)\)\), machine learning \(XGBoost\([Chen and Guestrin, 2016](https://arxiv.org/html/2611.22133#bib.bib9)\), LightGBM\([Ke et al\., 2017](https://arxiv.org/html/2611.22133#bib.bib29)\)\), deep learning \(FC\-LSTM\([Hochreiter and Schmidhuber, 1997](https://arxiv.org/html/2611.22133#bib.bib18)\), TCN\([Bai et al\., 2018](https://arxiv.org/html/2611.22133#bib.bib4)\), DLinear\([Zeng et al\., 2023](https://arxiv.org/html/2611.22133#bib.bib76)\), Informer\([Zhou et al\., 2021](https://arxiv.org/html/2611.22133#bib.bib77)\), PatchTST\([Nie et al\., 2023](https://arxiv.org/html/2611.22133#bib.bib46)\), iTransformer\([Liu et al\., 2024](https://arxiv.org/html/2611.22133#bib.bib37)\), DUET\([Qiu et al\., 2025](https://arxiv.org/html/2611.22133#bib.bib53)\), Vanilla MoE\), and a pretrained time\-series foundation model \(Chronos\-2\([Ansari et al\., 2024](https://arxiv.org/html/2611.22133#bib.bib2)\)\)\. For fair comparison, all models use the same time windows and the same inputs, including the observed 15\-day mobility sequence, the contextual variables, and the future STP\. ReMoE’s 16 router features are computed from these same inputs\. Our code repository details the training settings and hyperparameter configurations of all models\.

### 4\.2\.Main Performance \(EQ1\)

Table 1\.Model performance under in\-distribution \(ID\), leave\-one\-event\-out \(LOEO\), and leave\-one\-hazard\-type\-out \(LOHTO\) settings\. Best results are in bold and second\-best results are underlined\. Event\-level results are reported in Appendix\.SettingIn\-distributionLOEOLOHTOMetric groupPrediction accuracyRecovery semanticsPrediction accuracyRecovery semanticsPrediction accuracyRecovery semanticsMethodRMSE↓\\downarrowMAE↓\\downarrowR2↑R^\{2\}\\uparrowCDE↓\\downarrowRTE↓\\downarrowRMSE↓\\downarrowMAE↓\\downarrowR2↑R^\{2\}\\uparrowCDE↓\\downarrowRTE↓\\downarrowRMSE↓\\downarrowMAE↓\\downarrowR2↑R^\{2\}\\uparrowCDE↓\\downarrowRTE↓\\downarrowARIMAX0\.0990\.0780\.4230\.3561\.3550\.1120\.0860\.3070\.4781\.3420\.1280\.0980\.2290\.5791\.557XGBoost0\.0670\.0510\.7370\.2101\.0390\.0890\.0710\.5480\.3781\.1750\.1030\.0820\.4790\.4611\.370LightGBM0\.0640\.0480\.7610\.1970\.9890\.0870\.0690\.5670\.3661\.1450\.1010\.0800\.4950\.4481\.344DLinear0\.0730\.0510\.7040\.2161\.0520\.0880\.0690\.5550\.3751\.1270\.1020\.0790\.5010\.4421\.315FC\-LSTM0\.0660\.0490\.7420\.2051\.0350\.0840\.0650\.5940\.3511\.0950\.0990\.0780\.5250\.4261\.289Informer0\.0660\.0480\.7440\.1971\.0250\.0910\.0720\.5050\.3951\.1970\.1090\.0860\.4130\.4861\.414iTransformer0\.0570\.0410\.8120\.1650\.9630\.0850\.0670\.5880\.3571\.1140\.1020\.0810\.4850\.4441\.325PatchTST0\.0620\.0450\.7760\.1871\.0190\.0840\.0640\.6280\.3261\.0860\.0950\.0780\.5450\.4121\.302TCN0\.0580\.0420\.8030\.1700\.9820\.0790\.0620\.6350\.3331\.0650\.0970\.0760\.5650\.4211\.279DUET0\.0590\.0440\.7970\.2121\.0480\.0790\.0590\.6440\.3041\.1100\.0910\.0710\.5680\.4051\.275Vanilla MoE0\.0680\.0520\.7400\.2341\.0110\.0870\.0660\.6050\.3261\.0830\.0950\.0750\.5300\.4381\.290Chronos\-20\.0690\.0490\.7330\.2151\.0670\.0840\.0680\.5890\.3681\.1300\.1020\.0800\.4860\.4551\.352ReMoE0\.055\(\+3\.5%\)0\.042\(\-2\.4%\)0\.826\(\+0\.014\)0\.156\(\+5\.5%\)0\.975\(\-1\.2%\)0\.073\(\+7\.6%\)0\.058\(\+1\.7%\)0\.675\(\+0\.031\)0\.286\(\+5\.9%\)0\.983\(\+7\.7%\)0\.083\(\+8\.8%\)0\.069\(\+2\.8%\)0\.625\(\+0\.057\)0\.335\(\+17\.3%\)1\.082\(\+15\.1%\)w/o router features0\.0580\.0430\.8070\.1711\.0070\.0770\.0610\.6450\.3111\.0350\.0930\.0740\.5690\.3931\.236w/o MoE0\.0610\.0460\.7910\.1821\.0410\.0830\.0650\.5950\.3441\.1250\.1010\.0790\.5150\.4251\.269w/o STP\-RC0\.0640\.0490\.7850\.1911\.0250\.0790\.0620\.6310\.3521\.0750\.0940\.0750\.5550\.4471\.214

#### 4\.2\.1\.In\-distribution performance\.

We first evaluate ReMoE in the in\-distribution setting\. These results establish a baseline for assessing its subsequent OOD generalization capability\. As shown in Table[1](https://arxiv.org/html/2611.22133#S4.T1), ReMoE achieved the best results on three of the five metrics, with an RMSE of0\.0550\.055, anR2R^\{2\}of0\.8260\.826, and a CDE of0\.1560\.156, reducing RMSE and CDE by3\.5%3\.5\\%and5\.5%5\.5\\%, respectively, while improvingR2R^\{2\}by0\.0140\.014compared with the strongest baseline\. Its MAE of0\.0420\.042matched that of TCN and was only0\.0010\.001higher than iTransformer, while its RTE of0\.9750\.975was within0\.0120\.012days of the best\-performing model\. Notably, although ReMoE is designed to improve cross\-disaster generalization, it also shows competitive results under in\-distribution conditions\.

#### 4\.2\.2\.Cross\-disaster generalization\.

ReMoE’s advantage became more pronounced as the distribution shift increased\. Under LOEO, which evaluates transfer to unseen events of the same hazard types, ReMoE ranked first across all five metrics, achieving an RMSE of0\.0730\.073, an MAE of0\.0580\.058, and anR2R^\{2\}of0\.6750\.675\. Compared with the best baseline, it reduces errors by up to7\.7%7\.7\\%and improvesR2R^\{2\}by0\.0310\.031\. Under the more challenging LOHTO setting, ReMoE maintained the best performance across all metrics, reducing RMSE and MAE by8\.8%8\.8\\%and2\.8%2\.8\\%and improvingR2R^\{2\}by0\.0570\.057over the strongest baselines\. Interestingly, the gains are particularly evident in the recovery\-semantic metrics, where ReMoE reduces CDE and RTE by17\.3%17\.3\\%and15\.1%15\.1\\%, respectively\. Among the baselines, DUET generally performs best under both OOD settings, while TCN is also competitive under LOEO\. The widening gap from LOEO to LOHTO suggests that the recovery patterns ReMoE learns transfer better to unseen hazard types than those of generic forecasters\.

#### 4\.2\.3\.Ablation study\.

Ablation results are reported in Table[1](https://arxiv.org/html/2611.22133#S4.T1)\. We evaluate three variants:w/o router features, which replaces the 11 engineered mobility features with a learned encoding of the observed mobility sequence, while retaining the contextual features;w/o MoE, which replaces the five trend\-residual experts with a single MLP; andw/o STP\-RC, which directly uses theBaseas the final forecast\. Our results show that: \(1\) removing the router features causes the mildest degradation across all settings; \(2\) replacing MoE with a single MLP causes substantially greater degradation, with the gap widening under stronger distribution shifts, highlighting the importance of multi\-expert composition for generalization; and \(3\) removing STP\-RC has the largest impact under ID, increasing RMSE from0\.0550\.055to0\.0640\.064and reducingR2R^\{2\}from0\.8260\.826to0\.7850\.785, but its contribution grows less consistently under OOD settings\. This may be because the relationship between hazard pressure and mobility correction varies across unseen disasters, making it harder for STP\-RC to learn a stable mapping\.

### 4\.3\.Model Interpretability \(EQ2\)

![Six-panel interpretability analysis showing empirical recovery types, router-weight representations, recovery-type specialization, feature--routing associations, cross-hazard expert reuse, and spatial heterogeneity in dominant experts.](https://arxiv.org/html/2611.22133v1/figures/fig_interpretability.png)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\.Six\-panel interpretability analysis showing empirical recovery types, router\-weight representations, recovery\-type specialization, feature–routing associations, cross\-hazard expert reuse, and spatial heterogeneity in dominant experts\.#### 4\.3\.1\.Understanding Learned Expert Specialization\.

Here, we examine how expert specialization is associated with distinct recovery types\. For post hoc validation, we cluster complete recovery curves from 18 disaster events and identified four representative types as shown in Figure[3](https://arxiv.org/html/2611.22133#acmlabel2)\(a\)\. Type I shows mild impact and rapid recovery; Type II captures prolonged disruption; Type III exhibits bimodal recovery after a pronounced impact; and Type IV represents severe impact with slow, incomplete recovery\. Each sample was assigned to the nearest cluster centroid, while poorly matched samples were categorized as “Other\.” Detailed clustering procedures and recovery\-type descriptions are provided in the Appendix[B\.2](https://arxiv.org/html/2611.22133#A2.SS2)\.

In Figure[3](https://arxiv.org/html/2611.22133#acmlabel2)\(b\), we project the final learned routing\-weight vectors using t\-SNE and color each sample by its post\-hoc recovery type\. The result exhibits recognizable clusters, indicating that similar recovery types are associated with similar expert\-weight combinations\. A t\-SNE comparison of the routing spaces learned by a standard MoE and ReMoE is provided in the Appendix[D\.2](https://arxiv.org/html/2611.22133#A4.SS2)\. Building on this observation, we then quantify the correspondence between recovery types and experts by calculating the mean expert weight within each type, as shown in Figure[3](https://arxiv.org/html/2611.22133#acmlabel2)\(c\)\. It is observed that Type I samples are predominantly associated with E1 \(53\.7%53\.7\\%\), while Types II, III, and IV receive their highest weights for E2 \(38\.6%38\.6\\%\), E4 \(47\.1%47\.1\\%\), and E3 \(51\.2%51\.2\\%\), respectively\. E5 maintains the highest average weight for “other” \(40\.7%40\.7\\%\), potentially reflecting its statistical association with irregular recovery behaviors\. These associations provide evidence that our framework effectively learns specialized experts aligned with distinct recovery dynamics\.

#### 4\.3\.2\.Interpreting Expert Routing through Features\.

Another source of ReMoE’s interpretability comes from post hoc analysis of the explicit routing features, which reveals how input characteristics influence expert selection\. Figure[3](https://arxiv.org/html/2611.22133#acmlabel2)\(d\) presents the Spearman correlations between routing features and expert weights\. The lower part shows that the experts exhibit distinct feature\-association profiles\. E1 favors less\-impacted samples, as indicated by its negative association with impact depth\. E2 and E3 respond more to persistent or severe disruption, with E3 also associated with higher observed STP\. E4 is primarily related to recovery rate and momentum, whereas E5 is associated with greater persistence and recovery deficit but a lower restored ratio\. The upper part shows correlations among routing features, providing context for their associations with expert weights\.

We further conducted a global SHAP analysis to capture nonlinear feature contributions to router activations\. As reported in the Appendix[D\.3](https://arxiv.org/html/2611.22133#A4.SS3), impact depth, momentum, and builtup share dominate most to router activation\. Together, these results provide an input\-level post hoc explanation of the routing decisions\.

#### 4\.3\.3\.Semantic Characterization of Expert Composition\.

The third layer of interpretability arises from a post hoc characterization of how learned expert compositions vary across disasters and geographic space\. Figure[3](https://arxiv.org/html/2611.22133#acmlabel2)\(e\) shows distinct yet overlapping routing patterns across the four hazard types\. Hurricanes, for example, assign greater weights to E1 and E4, whereas floods place more emphasis on E2\. Given the associations between experts and recovery types, these patterns suggest that hurricanes more frequently combine dynamics associated with mild disruption and faster recovery, whereas floods emphasize prolonged disruption\. These patterns echo previous empirical findings\([Rajput and Mostafavi, 2023](https://arxiv.org/html/2611.22133#bib.bib56)\)\.

Figure[3](https://arxiv.org/html/2611.22133#acmlabel2)\(f\) maps each tract by its dominant expert, revealing substantial within\-event heterogeneity\. For instance, most tracts affected by the Vermont flood are dominated by E2, whereas a few activate E5, which receives the largest weight among samples outside the four types\. These maps can therefore flag localized anomalies for ad hoc response\. Moreover, unlike standalone post hoc analyses\([Hong et al\., 2021a](https://arxiv.org/html/2611.22133#bib.bib19);[Tang et al\., 2023](https://arxiv.org/html/2611.22133#bib.bib66)\), ReMoE has the potential to provide comparable prospective spatial characterizations across disasters\.

### 4\.4\.Robustness and Sensitivity Analysis \(EQ3\)

![Five-panel robustness analysis comparing future STP conditions, training-event availability, numbers of experts, prediction horizons, and post-event observation endpoints using RMSE.](https://arxiv.org/html/2611.22133v1/figures/fig_uncertainty.png)Figure 4\.Robustness and sensitivity analysis\.\(a\) RMSE by forecast lead under no\-future, forecast, and oracle STP\. \(b\) ReMoE and iTransformer across training\-event access levels\. \(c\)–\(e\) Sensitivity to expert count, prediction horizon, and observation endpoint\. All experiments use the ID setting\.Five\-panel robustness analysis comparing future STP conditions, training\-event availability, numbers of experts, prediction horizons, and post\-event observation endpoints using RMSE\.Under realistic deployment, practical conditions may vary\. We therefore evaluate ReMoE’s robustness and sensitivity\. First, we examine uncertainty in future hazard conditions\. In Figure[4](https://arxiv.org/html/2611.22133#acmlabel3)\(a\), Forecast STP consistently achieves lower errors than the no\-future setting and approaches Oracle STP performance, highlighting the value of forecast hazard information\. The increasing gap between Forecast and Oracle STP at longer leads may be due to the reduced reliability of hazard forecasts over time\. Figure[4](https://arxiv.org/html/2611.22133#acmlabel3)\(b\) evaluates ReMoE performance under limited training data\. Compared with the strongest baseline under the ID setting \(i\.e\., iTransformer\), ReMoE exhibits less performance degradation as less training data are available\. Its advantage is most pronounced with only20%20\\%of the training data\.

We further assess sensitivity to model configurations\. In Figure[4](https://arxiv.org/html/2611.22133#acmlabel3)\(c\), ReMoE remains stable across different numbers of experts when Top\-K is fixed at 2, with five experts achieving the lowest mean RMSE and variability\. We therefore adopt five experts as a balance between performance and interpretability\. In addition, different applications require different prediction horizons\. In practice, longer horizons are desirable if accuracy can be maintained\. As shown in Figure[4](https://arxiv.org/html/2611.22133#acmlabel3)\(d\), performance degrades moderately from 3 to 7 days but declines substantially at 14 days; thus, we select a 7\-day horizon\. Finally, extending the observation endpoint fromt0\+3t\_\{0\}\+3tot0\+5t\_\{0\}\+5improves performance, whereas later observations provide limited additional gains, as shown in Figure[4](https://arxiv.org/html/2611.22133#acmlabel3)\(e\)\. We therefore chooset0\+5t\_\{0\}\+5as the earliest endpoint achieving stable accuracy\.

## 5\.Discussion and Conclusion

##### Potential Applications\.

Disasters are sudden and may strike cities with limited prior experience of a particular hazard\. ReMoE provides a cross\-disaster prediction framework, that is to say, our predicted post\-disaster mobility patterns can be used as a reference for policymakers to identify fragile regions even under unseen hazards, and decide where to assess first\([Yabe et al\., 2020](https://arxiv.org/html/2611.22133#bib.bib72)\)\. This work further motivates future research on resilience curve prediction in other community systems, such as infrastructure networks\([Cao et al\., 2022](https://arxiv.org/html/2611.22133#bib.bib6)\)\.

From a broader AI\-for\-disaster perspective, we highlight the value of integrating scientific process knowledge with data\-driven learning\. Our results further suggest that model capacity per se may not be a “silver bullet” for complex disaster\-forecasting tasks such as ours: generic deep forecasters degrade under out\-of\-distribution conditions, while black\-box predictions remain difficult to interpret in high\-stakes decision contexts\([Rudin, 2019](https://arxiv.org/html/2611.22133#bib.bib58)\)\. ReMoE addresses these challenges through a feature\-guided MoE architecture whose experts are associated with distinct recovery types, providing an interpretable prediction pathway\. These traceable forecasts can support anticipatory assessment, although resource decisions still require field information\.

##### Conclusion\.

This paper introduces an interpretable mixture\-of\-experts framework \(namely, ReMoE\) to forecast human mobility recovery\. ReMoE is designed to generalize to unseen disasters while providing model interpretability\. To this end, we devise an IRTC module that integrates multiple specialized experts through weights learned from feature\-guided sparse router to compose heterogeneous recovery types\. IRTC outputs a basic recovery pattern that is forwarded into STP\-RC module, enabling the prediction to adapt to evolving future hazard conditions\. Extensive experiments on a large\-scale multi\-disaster mobility dataset demonstrate ReMoE’s superior forecasting performance, particularly for unseen disasters, while interpretability analyses show transparent routing behavior and expert specialization associated with recovery types\.

## 6\.Limitations and Ethical Considerations

Our study shares several limitations with prior research using large\-scale mobility data\([Li et al\., 2022](https://arxiv.org/html/2611.22133#bib.bib35);[Hsu et al\., 2024](https://arxiv.org/html/2611.22133#bib.bib21)\)\. First, the visit records cover only a subset of the population, and the data are aggregated at the census tract level, limiting analyses at smaller spatial units\. Second, our 18 events were selected for clear timelines, so more disaster cases, including compound events, are needed to draw robust conclusions\. Third, while we applied a consistent observation window \(t0−9t\_\{0\}\-9tot0\+5t\_\{0\}\+5\) across all events, future research should explore more context\-specific definitions of this window\. Finally, associations among router features, expert weights, and empirical recovery types can only be interpreted as statistical rather than causal evidence\.

## 7\.Generative AI Usage

To OpenAI Codex, for assisting with experimental implementation and language polishing\.

## References

- Ansari et al\.\(2024\)Abdul Fatir Ansari, Lorenzo Stella, Ali Caner Turkmen, Xiyuan Zhang, Pedro Mercado, Huibin Shen, Oleksandr Shchur, Syama Sundar Rangapuram, Sebastian Pineda Arango, Shubham Kapoor, Jasper Zschiegner, Danielle C\. Maddix, Hao Wang, Michael W\. Mahoney, Kari Torkkola, Andrew Gordon Wilson, Michael Bohlke\-Schneider, and Bernie Wang\. 2024\.Chronos: Learning the Language of Time Series\.*Transactions on Machine Learning Research*\(2024\)\.Expert Certification\.[https://openreview\.net/forum?id=gerNCVqqtR](https://openreview.net/forum?id=gerNCVqqtR)
- Anyidoho et al\.\(2023\)Prosper K\. Anyidoho, Xinglong Ju, Rachel A\. Davidson, and Linda K\. Nozick\. 2023\.A Machine Learning Approach for Predicting Hurricane Evacuee Destination Location Using Smartphone Location Data\.*Computational Urban Science*3, 1, Article 30 \(Sept\. 2023\)\.[doi:10\.1007/s43762\-023\-00102\-0](https://doi.org/10.1007/s43762-023-00102-0)
- Bai et al\.\(2018\)Shaojie Bai, J Zico Kolter, and Vladlen Koltun\. 2018\.An empirical evaluation of generic convolutional and recurrent networks for sequence modeling\.*arXiv preprint arXiv:1803\.01271*\(2018\)\.
- Box et al\.\(2015\)George EP Box, Gwilym M Jenkins, Gregory C Reinsel, and Greta M Ljung\. 2015\.*Time series analysis: forecasting and control*\.John Wiley & Sons\.
- Cao et al\.\(2022\)Quoc Dung Cao, Scott B\. Miles, and Youngjun Choe\. 2022\.Infrastructure recovery curve estimation using Gaussian process regression on expert elicited data\.*Reliability Engineering & System Safety*217 \(2022\), 108054\.[doi:10\.1016/j\.ress\.2021\.108054](https://doi.org/10.1016/j.ress.2021.108054)
- CDC/ATSDR \(2018\)CDC/ATSDR\. 2018\.Social Vulnerability Index 2018 Database, United States\.[https://www\.atsdr\.cdc\.gov/placeandhealth/svi/data\_documentation\_download\.html](https://www.atsdr.cdc.gov/placeandhealth/svi/data_documentation_download.html)\.Accessed: 2026\-05\-01\.
- Chapman et al\.\(2021\)Jim Chapman, Eric H Fox, William Bachman, Lawrence D Frank, John Thomas, and A Rourk Reyes\. 2021\.Smart location database\.*Environmental Protection Agency*\(2021\)\.
- Chen and Guestrin \(2016\)Tianqi Chen and Carlos Guestrin\. 2016\.XGBoost: A Scalable Tree Boosting System\. In*Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining*\(San Francisco, California, USA\)*\(KDD ’16\)*\. Association for Computing Machinery, New York, NY, USA, 785–794\.[doi:10\.1145/2939672\.2939785](https://doi.org/10.1145/2939672.2939785)
- Chen and He \(2021\)Xinlei Chen and Kaiming He\. 2021\.Exploring simple siamese representation learning\. In*Proceedings of the IEEE/CVF conference on computer vision and pattern recognition*\. 15750–15758\.
- Chikaraishi et al\.\(2020\)Makoto Chikaraishi, Prateek Garg, Varun Varghese, Kazuki Yoshizoe, Junji Urata, Yasuhiro Shiomi, and Ryuki Watanabe\. 2020\.On the possibility of short\-term traffic prediction during disaster with machine learning approaches: An exploratory analysis\.*Transport Policy*98 \(2020\), 91–104\.[doi:10\.1016/j\.tranpol\.2020\.05\.023](https://doi.org/10.1016/j.tranpol.2020.05.023)
- Donovan and Work \(2017\)Brian Donovan and Daniel B\. Work\. 2017\.Empirically quantifying city\-scale transportation system resilience to extreme events\.*Transportation Research Part C: Emerging Technologies*79 \(2017\), 333–346\.[doi:10\.1016/j\.trc\.2017\.03\.002](https://doi.org/10.1016/j.trc.2017.03.002)
- Fan et al\.\(2018\)Zipei Fan, Xuan Song, Tianqi Xia, Renhe Jiang, Ryosuke Shibasaki, and Ritsu Sakuramachi\. 2018\.Online Deep Ensemble Learning for Predicting Citywide Human Mobility\.*Proc\. ACM Interact\. Mob\. Wearable Ubiquitous Technol\.*2, 3, Article 105 \(Sept\. 2018\), 21 pages\.[doi:10\.1145/3264915](https://doi.org/10.1145/3264915)
- Federal Emergency Management Agency \(nd\)Federal Emergency Management Agency\. n\.d\.\.OpenFEMA Data Sets\.[https://www\.fema\.gov/about/openfema/data\-sets](https://www.fema.gov/about/openfema/data-sets)\.Accessed: 2026\-05\-22\.
- Guo et al\.\(2026\)Hongcan Guo, Haolang Lu, Guoshun Nan, Bolun Chu, Jialin Zhuang, Yuan Yang, Wenhao Che, Xinye Cao, Sicong Leng, Qimei Cui, et al\.2026\.Advancing expert specialization for better moe\.*Advances in Neural Information Processing Systems*38 \(2026\), 48767–48809\.
- Hao and Wang \(2023\)Haiyan Hao and Yan Wang\. 2023\.Modeling Dynamics of Community Resilience to Extreme Events with Explainable Deep Learning\.*Natural Hazards Review*24, 2 \(2023\), 04023013\.[doi:10\.1061/NHREFO\.NHENG\-1696](https://doi.org/10.1061/NHREFO.NHENG-1696)
- Hillebrand et al\.\(2018\)Helmut Hillebrand, Silke Langenheder, Karen Lebret, Eva Lindström, Örjan Östman, and Maren Striebel\. 2018\.Decomposing multiple dimensions of stability in global change experiments\.*Ecology Letters*21, 1 \(2018\), 21–30\.arXiv:https://onlinelibrary\.wiley\.com/doi/pdf/10\.1111/ele\.12867[doi:10\.1111/ele\.12867](https://doi.org/10.1111/ele.12867)
- Hochreiter and Schmidhuber \(1997\)Sepp Hochreiter and Jürgen Schmidhuber\. 1997\.Long short\-term memory\.*Neural computation*9, 8 \(1997\), 1735–1780\.
- Hong et al\.\(2021a\)Boyeong Hong, Bartosz J\. Bonczak, Arpit Gupta, and Constantine E\. Kontokosta\. 2021a\.Measuring Inequality in Community Resilience to Natural Disasters Using Large\-Scale Mobility Data\.*Nature Communications*12, 1, Article 1870 \(March 2021\)\.[doi:10\.1038/s41467\-021\-22160\-w](https://doi.org/10.1038/s41467-021-22160-w)
- Hong et al\.\(2021b\)Boyeong Hong, Bartosz J\. Bonczak, Arpit Gupta, and Constantine E\. Kontokosta\. 2021b\.Measuring inequality in community resilience to natural disasters using large\-scale mobility data\.*Nature Communications*12, 1 \(2021\), 1870\.[doi:10\.1038/s41467\-021\-22160\-w](https://doi.org/10.1038/s41467-021-22160-w)
- Hsu et al\.\(2024\)Chia\-Wei Hsu, Chenyue Liu, Kiet Minh Nguyen, Yu\-Heng Chien, and Ali Mostafavi\. 2024\.Do human mobility network analyses produced from different location\-based data sources yield similar results across scales?*Computers, Environment and Urban Systems*107 \(2024\), 102052\.[doi:10\.1016/j\.compenvurbsys\.2023\.102052](https://doi.org/10.1016/j.compenvurbsys.2023.102052)
- Hsu and Mostafavi \(2025\)Chia\-Wei Hsu and Ali Mostafavi\. 2025\.Dissecting Resilience Curve Archetypes and Properties in Human Systems Facing Weather Hazards\.*Scientific Reports*15, 1, Article 11897 \(April 2025\)\.[doi:10\.1038/s41598\-025\-95909\-8](https://doi.org/10.1038/s41598-025-95909-8)
- Huang et al\.\(2026\)Xinghao Huang, Shengyu Tao, Chen Liang, Yining Tang, Jiawei Chen, Junzhe Shi, Yuqi Li, Bizhong Xia, Guangmin Zhou, and Xuan Zhang\. 2026\.iMOE: prediction of second\-life battery degradation trajectory using interpretable mixture of experts\.*Nature Communications*17, 1 \(Feb\. 2026\), 2549\.[doi:10\.1038/s41467\-026\-69369\-1](https://doi.org/10.1038/s41467-026-69369-1)
- Jiang et al\.\(2017\)Fang\-Zhou Jiang, Lei Zhong, Kanchana Thilakarathna, Aruna Seneviratne, Kiyoshi Takano, Shigeki Yamada, and Yusheng Ji\. 2017\.Supercharging Crowd Dynamics Estimation in Disasters via Spatio\-Temporal Deep Neural Network\. In*2017 IEEE International Conference on Data Science and Advanced Analytics \(DSAA\)*\. 184–192\.[doi:10\.1109/DSAA\.2017\.11](https://doi.org/10.1109/DSAA.2017.11)
- Jiang et al\.\(2022\)Renhe Jiang, Zekun Cai, Zhaonan Wang, Chuang Yang, Zipei Fan, Quanjun Chen, Xuan Song, and Ryosuke Shibasaki\. 2022\.Predicting Citywide Crowd Dynamics at Big Events: A Deep Learning System\.*ACM Trans\. Intell\. Syst\. Technol\.*13, 2, Article 21 \(March 2022\), 24 pages\.[doi:10\.1145/3472300](https://doi.org/10.1145/3472300)
- Jiang et al\.\(2023a\)Renhe Jiang, Zhaonan Wang, Yudong Tao, Chuang Yang, Xuan Song, Ryosuke Shibasaki, Shu\-Ching Chen, and Mei\-Ling Shyu\. 2023a\.Learning Social Meta\-knowledge for Nowcasting Human Mobility in Disaster\. In*Proceedings of the ACM Web Conference 2023*\(Austin, TX, USA\)*\(WWW ’23\)*\. Association for Computing Machinery, New York, NY, USA, 2655–2665\.[doi:10\.1145/3543507\.3583991](https://doi.org/10.1145/3543507.3583991)
- Jiang et al\.\(2023b\)Yuqin Jiang, Faxi Yuan, Hamed Farahmand, Kushal Acharya, Jingdi Zhang, and Ali Mostafavi\. 2023b\.Data\-driven tracking of the bounce\-back path after disasters: Critical milestones of population activity recovery and their spatial inequality\.*International Journal of Disaster Risk Reduction*92 \(2023\), 103693\.[doi:10\.1016/j\.ijdrr\.2023\.103693](https://doi.org/10.1016/j.ijdrr.2023.103693)
- Karniadakis et al\.\(2021\)George Em Karniadakis, Ioannis G\. Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang\. 2021\.Physics\-informed machine learning\.*Nature Reviews Physics*3, 6 \(June 2021\), 422–440\.[doi:10\.1038/s42254\-021\-00314\-5](https://doi.org/10.1038/s42254-021-00314-5)
- Ke et al\.\(2017\)Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie\-Yan Liu\. 2017\.LightGBM: A Highly Efficient Gradient Boosting Decision Tree\. In*Advances in Neural Information Processing Systems*, I\. Guyon, U\. Von Luxburg, S\. Bengio, H\. Wallach, R\. Fergus, S\. Vishwanathan, and R\. Garnett \(Eds\.\), Vol\. 30\. Curran Associates, Inc\.[https://proceedings\.neurips\.cc/paper\_files/paper/2017/file/6449f44a102fde848669bdd9eb6b76fa\-Paper\.pdf](https://proceedings.neurips.cc/paper_files/paper/2017/file/6449f44a102fde848669bdd9eb6b76fa-Paper.pdf)
- Khaefi et al\.\(2018\)Muhammad Rizal Khaefi, Pamungkas Jutta Prahara, Muhammad Rheza, Dikara Alkarisya, and George Hodge\. 2018\.Predicting Evacuation Destinations due to a Natural Hazard using Mobile Network Data\. In*2018 2nd International Conference on Informatics and Computational Sciences \(ICICoS\)*\. 1–6\.[doi:10\.1109/ICICOS\.2018\.8621662](https://doi.org/10.1109/ICICOS.2018.8621662)
- Kieu et al\.\(2025\)Minh Kieu, Alexis Comber, Thanh Bui Quang, and Nick Malleson\. 2025\.Modelling social mobility disruptions and recovery during disasters: A mobile phone data approach\.*International Journal of Disaster Risk Reduction*130 \(2025\), 105812\.[doi:10\.1016/j\.ijdrr\.2025\.105812](https://doi.org/10.1016/j.ijdrr.2025.105812)
- Kokhlikyan et al\.\(2020\)Narine Kokhlikyan, Vivek Miglani, Miguel Martin, Edward Wang, Bilal Alsallakh, Jonathan Reynolds, Alexander Melnikov, Natalia Kliushkina, Carlos Araya, Siqi Yan, and Orion Reblitz\-Richardson\. 2020\.Captum: A unified and generic model interpretability library for PyTorch\.arXiv:2009\.07896 \[cs\.LG\][https://arxiv\.org/abs/2009\.07896](https://arxiv.org/abs/2009.07896)
- Li et al\.\(2024\)Jiahao Li, Huandong Wang, and Xinlei Chen\. 2024\.Physics\-informed Neural ODE for Post\-disaster Mobility Recovery\. In*Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining*\(Barcelona, Spain\)*\(KDD ’24\)*\. Association for Computing Machinery, New York, NY, USA, 1587–1598\.[doi:10\.1145/3637528\.3672027](https://doi.org/10.1145/3637528.3672027)
- Li et al\.\(2026\)Wenjing Li, Jinyu Chen, Yuhao Yao, Peiran Li, Hill Hiroki Kobayashi, Haoran Zhang, Xuan Song, Ryosuke Shibasaki, and Xiaodan Shi\. 2026\.Learn to Cluster Human Mobility Pattern for Post\-Disaster Analysis\.*IEEE Transactions on Big Data*12, 3 \(2026\), 936–947\.[doi:10\.1109/TBDATA\.2026\.3653656](https://doi.org/10.1109/TBDATA.2026.3653656)
- Li et al\.\(2022\)Weiyu Li, Qi Wang, Yuanyuan Liu, Mario L\. Small, and Jianxi Gao\. 2022\.A spatiotemporal decay model of human mobility when facing large\-scale crises\.*Proceedings of the National Academy of Sciences*119, 33 \(2022\), e2203042119\.[doi:10\.1073/pnas\.2203042119](https://doi.org/10.1073/pnas.2203042119)
- Lim et al\.\(2021\)Bryan Lim, Sercan Ö Arık, Nicolas Loeff, and Tomas Pfister\. 2021\.Temporal fusion transformers for interpretable multi\-horizon time series forecasting\.*International journal of forecasting*37, 4 \(2021\), 1748–1764\.
- Liu et al\.\(2024\)Yong Liu, Tengge Hu, Haoran Zhang, Haixu Wu, Shiyu Wang, Lintao Ma, and Mingsheng Long\. 2024\.iTransformer: Inverted Transformers Are Effective for Time Series Forecasting\. In*International Conference on Learning Representations*, B\. Kim, Y\. Yue, S\. Chaudhuri, K\. Fragkiadaki, M\. Khan, and Y\. Sun \(Eds\.\), Vol\. 2024\. 11116–11140\.[https://proceedings\.iclr\.cc/paper\_files/paper/2024/file/2ea18fdc667e0ef2ad82b2b4d65147ad\-Paper\-Conference\.pdf](https://proceedings.iclr.cc/paper_files/paper/2024/file/2ea18fdc667e0ef2ad82b2b4d65147ad-Paper-Conference.pdf)
- Lyu \(2025\)Yilin Lyu\. 2025\.Enhancing Approximate Conformance Checking Accuracy with Hierarchical Clustering Model Behaviour Sampling\.*International Journal of Advanced Computer Science and Applications*16, 8 \(2025\)\.[doi:10\.14569/IJACSA\.2025\.0160803](https://doi.org/10.14569/IJACSA.2025.0160803)
- Lyu et al\.\(2025\)Yilin Lyu, Abel Armas\-Cervantes, and Antonette Mendoza\. 2025\.Uncovering Self\-Regulated Learning Patterns in Chatbot\-Powered Environment through a Process Mining Lens\. In*36th Australasian Association for Engineering Education Annual Conference \(AAEE2025\): Engineering Education: Between the Human and the Digital*\. Engineers Australia, Brisbane, QLD, 9 pages\.[https://search\.informit\.org/doi/10\.3316/informit\.T2026031700007601807050328](https://search.informit.org/doi/10.3316/informit.T2026031700007601807050328)
- Lyu et al\.\(2026\)Yilin Lyu, Kang Liu, Zhongcai Cao, Yuxiao Luo, Ling Yin, Tengfei Yang, and Zugang Chen\. 2026\.FLAIR: An LLM\-integrated multi\-layer framework for fine\-grained activity type inference in human trajectory data\.*Transportation Research Part C: Emerging Technologies*190 \(2026\), 105787\.[doi:10\.1016/j\.trc\.2026\.105787](https://doi.org/10.1016/j.trc.2026.105787)
- National Interagency Fire Center \(nd\)National Interagency Fire Center\. n\.d\.\.Wildland Fire Open Data\.[https://data\-nifc\.opendata\.arcgis\.com/](https://data-nifc.opendata.arcgis.com/)\.Accessed: 2026\-05\-23\.
- National Oceanic and Atmospheric Administration \(\[n\. d\.\]\)National Oceanic and Atmospheric Administration\. \[n\. d\.\]\.Climate Data Online: Web Services Documentation\.[https://www\.ncdc\.noaa\.gov/cdo\-web/webservices/v2](https://www.ncdc.noaa.gov/cdo-web/webservices/v2)\.Accessed: 2026\-05\-01\.
- National Oceanic and Atmospheric Administration \(nda\)National Oceanic and Atmospheric Administration\. n\.d\.a\.NOAA High\-Resolution Rapid Refresh \(HRRR\) Model\.[https://registry\.opendata\.aws/noaa\-hrrr\-pds/](https://registry.opendata.aws/noaa-hrrr-pds/)\.Accessed: 2026\-05\-23\.
- National Oceanic and Atmospheric Administration \(ndb\)National Oceanic and Atmospheric Administration\. n\.d\.b\.NOAA National Water Model CONUS Retrospective Dataset\.[https://registry\.opendata\.aws/nwm\-archive/](https://registry.opendata.aws/nwm-archive/)\.Accessed: 2026\-05\-23\.
- National Weather Service \(nd\)National Weather Service\. n\.d\.\.National Weather Service\.[https://www\.weather\.gov/](https://www.weather.gov/)\.Accessed: 2026\-05\-02\.
- Nie et al\.\(2023\)Yuqi Nie, Nam H Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam\. 2023\.A Time Series is Worth 64 Words: Long\-term Forecasting with Transformers\. In*The Eleventh International Conference on Learning Representations*\.[https://openreview\.net/forum?id=Jbdc0vTOcol](https://openreview.net/forum?id=Jbdc0vTOcol)
- NOAA National Centers for Environmental Information \(nda\)NOAA National Centers for Environmental Information\. n\.d\.a\.Global Forecast System \(GFS\)\.[https://www\.ncei\.noaa\.gov/products/weather\-climate\-models/global\-forecast](https://www.ncei.noaa.gov/products/weather-climate-models/global-forecast)\.Accessed: 2026\-05\-23\.
- NOAA National Centers for Environmental Information \(ndb\)NOAA National Centers for Environmental Information\. n\.d\.b\.Global Historical Climatology Network Daily \(GHCN\-Daily\)\.[https://www\.ncei\.noaa\.gov/products/land\-based\-station/global\-historical\-climatology\-network\-daily](https://www.ncei.noaa.gov/products/land-based-station/global-historical-climatology-network-daily)\.Accessed: 2026\-05\-23\.
- NOAA National Hurricane Center \(2026\)NOAA National Hurricane Center\. 2026\.Atlantic Hurricane Database \(HURDAT2\), 1851–2025\.[https://www\.nhc\.noaa\.gov/data/hurdat/hurdat2\-1851\-2025\-02272026\.txt](https://www.nhc.noaa.gov/data/hurdat/hurdat2-1851-2025-02272026.txt)\.Updated February 27, 2026; accessed 2026\-05\-22\.
- NOAA National Hurricane Center \(nd\)NOAA National Hurricane Center\. n\.d\.\.Tropical Cyclone Reports\.[https://www\.nhc\.noaa\.gov/data/tcr/](https://www.nhc.noaa.gov/data/tcr/)\.Accessed: 2026\-05\-22\.
- NOAA Physical Sciences Laboratory \(nd\)NOAA Physical Sciences Laboratory\. n\.d\.\.Gridded Climate Data\.[https://psl\.noaa\.gov/data/gridded/index\.html](https://psl.noaa.gov/data/gridded/index.html)\.Accessed: 2026\-05\-20\.
- Poulin and Kane \(2021\)Craig Poulin and Michael B\. Kane\. 2021\.Infrastructure resilience curves: Performance measures and summary metrics\.*Reliability Engineering & System Safety*216 \(2021\), 107926\.[doi:10\.1016/j\.ress\.2021\.107926](https://doi.org/10.1016/j.ress.2021.107926)
- Qiu et al\.\(2025\)Xiangfei Qiu, Xingjian Wu, Yan Lin, Chenjuan Guo, Jilin Hu, and Bin Yang\. 2025\.DUET: Dual Clustering Enhanced Multivariate Time Series Forecasting\. In*Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V\.1*\(Toronto ON, Canada\)*\(KDD ’25\)*\. Association for Computing Machinery, New York, NY, USA, 1185–1196\.[doi:10\.1145/3690624\.3709325](https://doi.org/10.1145/3690624.3709325)
- Rachunok and Nateghi \(2021\)Benjamin Rachunok and Roshanak Nateghi\. 2021\.Overemphasis on Recovery Inhibits Community Transformation and Creates Resilience Traps\.*Nature Communications*12, 1 \(2021\), 7331\.[doi:10\.1038/s41467\-021\-27359\-5](https://doi.org/10.1038/s41467-021-27359-5)
- Rahman and Hasan \(2018\)Rezaur Rahman and Samiul Hasan\. 2018\.Short\-Term Traffic Speed Prediction for Freeways During Hurricane Evacuation: A Deep Learning Approach\. In*2018 21st International Conference on Intelligent Transportation Systems \(ITSC\)*\. 1291–1296\.[doi:10\.1109/ITSC\.2018\.8569443](https://doi.org/10.1109/ITSC.2018.8569443)
- Rajput and Mostafavi \(2023\)Akhil Anil Rajput and Ali Mostafavi\. 2023\.Latent sub\-structural resilience mechanisms in temporal human mobility networks during urban flooding\.*Scientific Reports*13, 1 \(July 2023\), 10953\.[doi:10\.1038/s41598\-023\-37965\-6](https://doi.org/10.1038/s41598-023-37965-6)
- Roy et al\.\(2021\)Kamol Chandra Roy, Samiul Hasan, Aron Culotta, and Naveen Eluru\. 2021\.Predicting traffic demand during hurricane evacuation using Real\-time data from transportation systems and social media\.*Transportation Research Part C: Emerging Technologies*131 \(2021\), 103339\.[doi:10\.1016/j\.trc\.2021\.103339](https://doi.org/10.1016/j.trc.2021.103339)
- Rudin \(2019\)Cynthia Rudin\. 2019\.Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and Use Interpretable Models Instead\.*Nature Machine Intelligence*1, 5 \(May 2019\), 206–215\.[doi:10\.1038/s42256\-019\-0048\-x](https://doi.org/10.1038/s42256-019-0048-x)
- Shazeer et al\.\(2017\)Noam Shazeer, \*Azalia Mirhoseini, \*Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean\. 2017\.Outrageously Large Neural Networks: The Sparsely\-Gated Mixture\-of\-Experts Layer\. In*International Conference on Learning Representations*\.[https://openreview\.net/forum?id=B1ckMDqlg](https://openreview.net/forum?id=B1ckMDqlg)
- Shimosaka et al\.\(2015\)Masamichi Shimosaka, Keisuke Maeda, Takeshi Tsukiji, and Kota Tsubouchi\. 2015\.Forecasting urban dynamics with mobility logs by bilinear Poisson regression\. In*Proceedings of the 2015 ACM International Joint Conference on Pervasive and Ubiquitous Computing*\(Osaka, Japan\)*\(UbiComp ’15\)*\. Association for Computing Machinery, New York, NY, USA, 535–546\.[doi:10\.1145/2750858\.2807527](https://doi.org/10.1145/2750858.2807527)
- Song et al\.\(2017\)Xuan Song, Ryosuke Shibasaki, Nicholos Yuan, Xing Xie, Tao Li, and Ryutaro Adachi\. 2017\.DeepMob: Learning Deep Knowledge of Human Emergency Behavior and Mobility from Big and Heterogeneous Data\.*ACM Transactions on Information Systems*35 \(06 2017\), 1–19\.[doi:10\.1145/3057280](https://doi.org/10.1145/3057280)
- Song et al\.\(2013\)Xuan Song, Quanshi Zhang, Yoshihide Sekimoto, Teerayut Horanont, Satoshi Ueyama, and Ryosuke Shibasaki\. 2013\.Modeling and probabilistic reasoning of population evacuation during large\-scale disaster\. In*Proceedings of the 19th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining*\(Chicago, Illinois, USA\)*\(KDD ’13\)*\. Association for Computing Machinery, New York, NY, USA, 1231–1239\.[doi:10\.1145/2487575\.2488189](https://doi.org/10.1145/2487575.2488189)
- Song et al\.\(2014\)Xuan Song, Quanshi Zhang, Yoshihide Sekimoto, and Ryosuke Shibasaki\. 2014\.Prediction of human emergency behavior and their mobility following large\-scale disaster\. In*Proceedings of the 20th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining*\(New York, New York, USA\)*\(KDD ’14\)*\. Association for Computing Machinery, New York, NY, USA, 5–14\.[doi:10\.1145/2623330\.2623628](https://doi.org/10.1145/2623330.2623628)
- Swamy et al\.\(2025\)Vinitra Swamy, Syrielle Montariol, Julian Blackwell, Jibril Frej, Martin Jaggi, and Tanja Käser\. 2025\.Intrinsic User\-Centric Interpretability through Global Mixture of Experts\.arXiv:2402\.02933 \[cs\.LG\][https://arxiv\.org/abs/2402\.02933](https://arxiv.org/abs/2402.02933)
- Tang et al\.\(2024\)Junqing Tang, Jing Wang, Jiaying Li, Pengjun Zhao, Wei Lyu, Wei Zhai, Li Yuan, Li Wan, and Chenyu Yang\. 2024\.Predicting human mobility flows in response to extreme urban floods: A hybrid deep learning model considering spatial heterogeneity\.*Computers, Environment and Urban Systems*113 \(2024\), 102160\.[doi:10\.1016/j\.compenvurbsys\.2024\.102160](https://doi.org/10.1016/j.compenvurbsys.2024.102160)
- Tang et al\.\(2023\)Junqing Tang, Pengjun Zhao, Zhaoya Gong, Hongbo Zhao, Fengjue Huang, Jiaying Li, Zhihe Chen, Ling Yu, and Jun Chen\. 2023\.Resilience patterns of human mobility in response to extreme urban floods\.*National Science Review*10, 8 \(08 2023\), nwad097\.arXiv:https://academic\.oup\.com/nsr/article\-pdf/10/8/nwad097/50733886/nwad097\.pdf[doi:10\.1093/nsr/nwad097](https://doi.org/10.1093/nsr/nwad097)
- Tang et al\.\(2026\)Yinzhou Tang, Huandong Wang, Xiaochen Fan, and Yong Li\. 2026\.Predicting Human Mobility during Extreme Events via LLM\-Enhanced Cross\-City Learning\.arXiv:2507\.19737 \[cs\.LG\][https://arxiv\.org/abs/2507\.19737](https://arxiv.org/abs/2507.19737)
- U\.S\. Geological Survey \(nd\)U\.S\. Geological Survey\. n\.d\.\.USGS Water Data for the Nation: National Water Information System\.[https://waterdata\.usgs\.gov/](https://waterdata.usgs.gov/)\.Accessed: 2026\-05\-24\.
- Wang et al\.\(2023\)Ruoxi Wang, Qi Wang, and Nan Li\. 2023\.Percolation transitions in urban mobility networks in America’s 50 largest cities\.*Sustainable Cities and Society*91 \(2023\), 104435\.[doi:10\.1016/j\.scs\.2023\.104435](https://doi.org/10.1016/j.scs.2023.104435)
- Wang and Zhu \(2025\)Sheng Wang and Di Zhu\. 2025\.Inferring human movements after snowfall: a weather\-informed graph learning model for flow redistribution in mobility networks\.*International Journal of Geographical Information Science*39, 11 \(2025\), 2570–2598\.[doi:10\.1080/13658816\.2025\.2524394](https://doi.org/10.1080/13658816.2025.2524394)
- Wang et al\.\(2022\)Zhaonan Wang, Renhe Jiang, Hao Xue, Flora D Salim, Xuan Song, and Ryosuke Shibasaki\. 2022\.Event\-Aware Multimodal Mobility Nowcasting\. In*Proceedings of the AAAI Conference on Artificial Intelligence*, Vol\. 36\. 4228–4236\.
- Yabe et al\.\(2020\)Takahiro Yabe, Kota Tsubouchi, Naoya Fujiwara, Yoshihide Sekimoto, and Satish V\. Ukkusuri\. 2020\.Understanding post\-disaster population recovery patterns\.*Journal of The Royal Society Interface*17, 163 \(02 2020\), 20190532\.[doi:10\.1098/rsif\.2019\.0532](https://doi.org/10.1098/rsif.2019.0532)
- Yabe et al\.\(2016\)Takahiro Yabe, Kota Tsubouchi, Akihito Sudo, and Yoshihide Sekimoto\. 2016\.Predicting irregular individual movement following frequent mid\-level disasters using location data from smartphones\. In*Proceedings of the 24th ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems*\(Burlingame, California\)*\(SIGSPACIAL ’16\)*\. Association for Computing Machinery, New York, NY, USA, Article 54, 4 pages\.[doi:10\.1145/2996913\.2996929](https://doi.org/10.1145/2996913.2996929)
- Yang et al\.\(2025a\)Xiaojie Yang, Hangli Ge, Jiawei Wang, Zipei Fan, Renhe Jiang, Ryosuke Shibasaki, and Noboru Koshizuka\. 2025a\.CausalMob: Causal Human Mobility Prediction with LLMs\-derived Human Intentions toward Public Events\. In*Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V\.1*\(Toronto ON, Canada\)*\(KDD ’25\)*\. Association for Computing Machinery, New York, NY, USA, 1773–1784\.[doi:10\.1145/3690624\.3709231](https://doi.org/10.1145/3690624.3709231)
- Yang et al\.\(2025b\)Xingyi Yang, Constantin Venhoff, Ashkan Khakzar, Christian Schroeder De Witt, Puneet K Dokania, Adel Bibi, and Philip Torr\. 2025b\.Mixture of experts made intrinsically interpretable\.*arXiv preprint arXiv:2503\.07639*\(2025\)\.
- Zeng et al\.\(2023\)Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu\. 2023\.Are Transformers Effective for Time Series Forecasting?*Proceedings of the AAAI Conference on Artificial Intelligence*37, 9 \(Jun\. 2023\), 11121–11128\.[doi:10\.1609/aaai\.v37i9\.26317](https://doi.org/10.1609/aaai.v37i9.26317)
- Zhou et al\.\(2021\)Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang\. 2021\.Informer: Beyond Efficient Transformer for Long Sequence Time\-Series Forecasting\.*Proceedings of the AAAI Conference on Artificial Intelligence*35, 12 \(May 2021\), 11106–11115\.[doi:10\.1609/aaai\.v35i12\.17325](https://doi.org/10.1609/aaai.v35i12.17325)

## Appendix ALiterature Review

Table 2\.Existing studies on post\-disaster mobility prediction, where T and ST denote temporal and spatio\-temporal models, respectively, Interp\. denotes model interpretability, and Cross denotes cross\-disaster generalization\.LiteratureTargetT/STMethodFeaturesSpatial unitTemporal settingInterp\.CrossStatistical / probabilistic methods[Song et al\. \(2013\)](https://arxiv.org/html/2611.22133#bib.bib62)Evacuation mobilitySTMarkov decision processHistorical mobility, home locations, evacuation\-network attributes, travel attributes10 km grid12–81 d✓✗[Shimosaka et al\. \(2015\)](https://arxiv.org/html/2611.22133#bib.bib60)Active population countTPoisson regressionHistorical mobility, temporal covariates, weather covariates, location attributes900 m grid30–180 d✓✗[Song et al\. \(2014\)](https://arxiv.org/html/2611.22133#bib.bib63)Emergency behavior \(Individual\)STHMMHistorical mobility, important locations, disaster intensity, distance to hazard, government/news reports, population\-flow attributesRegional4 h–1 d✓✗[Kieu et al\. \(2025\)](https://arxiv.org/html/2611.22133#bib.bib31)POI visitsTHawkes processHistorical POI visits, POI categories, dwell time, temporal covariatesRegional1–3 h✓✗Machine\-learning methods[Khaefi et al\. \(2018\)](https://arxiv.org/html/2611.22133#bib.bib30)Evacuation destination \(Individual\)STXGBoostMobile\-phone usage, historical mobility, mobility regularity, and mobility diversityRegionalEvent period✗✗[Chikaraishi et al\. \(2020\)](https://arxiv.org/html/2611.22133#bib.bib11)Traffic flowTRF/SVM/XGBoostHistorical traffic states, occupancy, and temporal covariatesRoad corridor5–20 min✓✗[Anyidoho et al\. \(2023\)](https://arxiv.org/html/2611.22133#bib.bib3)Evacuation OD flowSTXGBoostOD distance, population, and hazard attributesRegionalEvent period✗✗Deep\-learning methods[Song et al\. \(2017\)](https://arxiv.org/html/2611.22133#bib.bib61)Evacuation mobility \(Individual\)STDBNHistorical mobility trajectories, hazard attributes, transportation\-network attributes, and news signalsCounty6 h✓✗[Jiang et al\. \(2017\)](https://arxiv.org/html/2611.22133#bib.bib24)Crowd distributionTLSTMHistorical crowd distributions and land\-use attributes500 m gridHourly✓✗[Rahman and Hasan \(2018\)](https://arxiv.org/html/2611.22133#bib.bib55)Traffic speedTLSTM\-NNHistorical traffic states from upstream, target, and downstream linksFreeway links5–15 min✗✗[Fan et al\. \(2018\)](https://arxiv.org/html/2611.22133#bib.bib13)Next location \(Individual\)TGRURecent location sequences and temporal covariates1 km grid15 min✗✗[Roy et al\. \(2021\)](https://arxiv.org/html/2611.22133#bib.bib57)Traffic volumeTLSTMHistorical traffic states, landfall timing, social\-media signals, evacuation orders, and weather covariatesEvacuation region1–30 h✗✗[Jiang et al\. \(2022\)](https://arxiv.org/html/2611.22133#bib.bib25)Crowd density/flowSTConvLSTMHistorical crowd\-density and crowd\-flow tensors500 m grid30–60 min✗✓[Wang et al\. \(2022\)](https://arxiv.org/html/2611.22133#bib.bib71)Mobility tensorSTGCRUHistorical multimodal mobility, temporal covariates, and adaptive graph edges1 km grid8 steps✗✓[Hao and Wang \(2023\)](https://arxiv.org/html/2611.22133#bib.bib16)Mobility frequencySTGCN\-LSTMHistorical mobility, social\-vulnerability attributes, built\-environment, and weather covariatesCensus tractDaily✓✗[Jiang et al\. \(2023a\)](https://arxiv.org/html/2611.22133#bib.bib26)Crowd flowSTGraph\-GRU \+ meta\-learningHistorical mobility, social\-media signals, spatial\-context attributes, and temporal covariatesCounty6 h/6 h✗✓[Tang et al\. \(2024\)](https://arxiv.org/html/2611.22133#bib.bib65)Mobility inflowSTGCN \+ Seq2Seq\-LSTMHistorical OD flows, adjacency structure, and network\-centrality attributesUrban grids6 h/1 h✗✗[Li et al\. \(2024\)](https://arxiv.org/html/2611.22133#bib.bib33)Mobility graphSTGraph Neural ODEInitial abnormal mobility graph, baseline mobility graph, and node/edge mobility attributesCountyDaily✗✗[Wang and Zhu \(2025\)](https://arxiv.org/html/2611.22133#bib.bib70)Mobility flow probabilitiesSTDynamic graph learningPre\-event mobility flows, node attributes, weather covariates, and edge attributesCensus tractDaily✗✗LLM\-enhanced methods[Yang et al\. \(2025a\)](https://arxiv.org/html/2611.22133#bib.bib74)Mobility volumeSTST\-ResNet/GCN \+ LLMHistorical mobility, POI attributes, temporal/region embeddings, OD graph, and LLM\-derived event\-intention featuresRegional1–24 h✗✓[Tang et al\. \(2026\)](https://arxiv.org/html/2611.22133#bib.bib67)Next location \(Individual\)STLLM\-enhanced hybrid modelCross\-city mobility trajectories, hazard attributes, and LLM\-derived intention featuresCandidate locationsNext step✗✗
Abbreviations:Markov decision process \(MDP\), hidden Markov model \(HMM\), random forest \(RF\), support vector machine \(SVM\), extreme gradient boosting \(XGBoost\), deep belief network \(DBN\), long short\-term memory \(LSTM\), gated recurrent unit \(GRU\), convolutional LSTM \(ConvLSTM\), graph convolutional network \(GCN\), graph convolutional recurrent unit \(GCRU\), neural ordinary differential equation \(Neural ODE\), spatio\-temporal residual network \(ST\-ResNet\)\.

### A\.1\.Post\-Disaster Human Mobility Forecasting

By post\-disaster human mobility, we refer to the short\-term disruption and recovery of human movement following a disaster\. Early studies mainly adopted probabilistic and statistical approaches\.[Song et al\. \(2013\)](https://arxiv.org/html/2611.22133#bib.bib62)modeled evacuation behavior as a Markov decision process based on an evacuation graph, while[Song et al\. \(2014\)](https://arxiv.org/html/2611.22133#bib.bib63)used a hidden Markov model\. Despite their interpretability, these methods relied on simplifying assumptions and were largely event\-specific\. Machine learning \(ML\) methods subsequently introduced models such as XGBoost to predict evacuation destinations and origin–destination flows using historical mobility, demographic, and hazard\-exposure features\([Anyidoho et al\., 2023](https://arxiv.org/html/2611.22133#bib.bib3);[Khaefi et al\., 2018](https://arxiv.org/html/2611.22133#bib.bib30)\)\. However, these approaches remained limited in capturing spatio\-temporal dependencies\.

DL methods were subsequently introduced to learn nonlinear mobility dynamics\. Early DL approaches mainly adopted recurrent architectures\. For instance,[Jiang et al\. \(2017\)](https://arxiv.org/html/2611.22133#bib.bib24)introduced an LSTM\-based model to learn temporal mobility patterns from sequential observations\. However, mobility dynamics are inherently spatio\-temporal, motivating models that jointly capture temporal evolution and spatial interactions\. Architectures such as ConvLSTM and GCN–LSTM combined recurrent networks with convolutional or graph\-based operators to learn regional dependencies and dynamic mobility patterns\([Jiang et al\., 2022](https://arxiv.org/html/2611.22133#bib.bib25);[Hao and Wang, 2023](https://arxiv.org/html/2611.22133#bib.bib16)\)\. Recent studies have incorporated adaptive graph learning and LLM\-derived external signals\([Jiang et al\., 2023a](https://arxiv.org/html/2611.22133#bib.bib26);[Yang et al\., 2025a](https://arxiv.org/html/2611.22133#bib.bib74)\), while Neural ODEs model mobility recovery as a continuous dynamic process\([Li et al\., 2024](https://arxiv.org/html/2611.22133#bib.bib33)\)\. However, these models remain difficult to interpret, and their spatial representations may overfit location\-specific dependencies, limiting transferability to unseen regions\.

### A\.2\.Interpretability in Post\-disaster Mobility Forecasting

Interpretability is essential for post\-disaster mobility forecasting, as transparent predictions can help policymakers understand mobility disruptions and design effective recovery strategies\. However, it remains underexplored in existing post\-disaster mobility forecasting studies\. Traditional statistical and ML models provide relatively direct interpretations through explicit coefficients, decision structures, or feature importance analyses\([Yabe et al\., 2016](https://arxiv.org/html/2611.22133#bib.bib73);[Shimosaka et al\., 2015](https://arxiv.org/html/2611.22133#bib.bib60);[Song et al\., 2013](https://arxiv.org/html/2611.22133#bib.bib62);[Chikaraishi et al\., 2020](https://arxiv.org/html/2611.22133#bib.bib11);[Jiang et al\., 2017](https://arxiv.org/html/2611.22133#bib.bib24);[Song et al\., 2017](https://arxiv.org/html/2611.22133#bib.bib61);[Hao and Wang, 2023](https://arxiv.org/html/2611.22133#bib.bib16)\)\. In contrast, DL models are more difficult to interpret due to their nonlinear architectures, with only limited efforts exploring post\-hoc explanations\. For example,[Jiang et al\. \(2017\)](https://arxiv.org/html/2611.22133#bib.bib24)examined temporal and spatial correlations to characterize learned mobility patterns, while[Hao and Wang \(2023\)](https://arxiv.org/html/2611.22133#bib.bib16)applied DeepLIFT\-based attribution to a GCN–LSTM model to quantify the contribution of community attributes to predicted mobility disruptions\. However, existing studies rarely go beyond post\-hoc, feature\-level explanations, leaving the intrinsic interpretability of forecasting models largely unexplored\([Yang et al\., 2025b](https://arxiv.org/html/2611.22133#bib.bib75);[Lyu et al\., 2025](https://arxiv.org/html/2611.22133#bib.bib39);[Lyu, 2025](https://arxiv.org/html/2611.22133#bib.bib38)\)\. ReMoE bridges this gap by combining both: its feature\-guided router makes expert allocation traceable, while post\-hoc analysis connects learned expert specialization with empirical recovery types, revealing the distinct semantic roles of different experts\.

### A\.3\.Cross\-Disaster Generalization

To support effective disaster management, mobility forecasting models must be capable of predicting responses to unprecedented events\([Lyu et al\., 2026](https://arxiv.org/html/2611.22133#bib.bib40)\)\. However, mobility responses vary across locations, hazard types, and severities, making models trained on specific cases difficult to generalize to unseen disasters\. Celebrated approaches have been explored to address this challenge, including transfer learning strategies that transfer mobility patterns across locations and reinforcement learning methods that represent disaster impacts through inferred reward functions\([Fan et al\., 2018](https://arxiv.org/html/2611.22133#bib.bib13);[Rahman and Hasan, 2018](https://arxiv.org/html/2611.22133#bib.bib55)\)\. Although these approaches are promising contributions, limitations remain in their generalizability\. For example, existing methods improve transferability by transferring observed patterns or learned representations rather than the underlying recovery dynamics shared across disasters\. Inspired by prior work on modeling cross\-disaster recovery dynamics\([Li et al\., 2022](https://arxiv.org/html/2611.22133#bib.bib35);[Yabe et al\., 2020](https://arxiv.org/html/2611.22133#bib.bib72);[Hsu and Mostafavi, 2025](https://arxiv.org/html/2611.22133#bib.bib22);[Tang et al\., 2023](https://arxiv.org/html/2611.22133#bib.bib66)\), ReMoE addresses this gap by learning reusable recovery types through expert specialization and enabling transferable prediction\.

## Appendix BDataset and Recovery Types

### B\.1\.Dataset Description

![A U.S. map showing 18 hurricanes, floods, winter storms,and wildfires, surrounded by their normalized tract-level mobilityrecovery curves.](https://arxiv.org/html/2611.22133v1/figures/disaster_study_areas_and_recovery_curves.png)Figure 5\.Disaster events and mobility recovery curves\.Event locations use marker size and color to indicate affected\-area extent and hazard severity\. Panels show normalized mobility: blue lines and shading indicate the mean and interquartile range of three\-day causally smoothed tract curves; gray dashed lines indicate the unsmoothed mean\. Red lines mark the event date\.A U\.S\. map showing 18 hurricanes, floods, winter storms,and wildfires, surrounded by their normalized tract\-level mobilityrecovery curves\.#### B\.1\.1\.Event Selection\.

Our multi\-disaster dataset spans different severity levels\. As shown in Fig\.[5](https://arxiv.org/html/2611.22133#acmlabel4), it contains 18 major U\.S\. disasters between 2019 and 2025, including eight hurricanes, four winter storms, two floods, and four wildfires\. Events were selected when \(i\) their chronology and principal impact area could be established from official records, \(ii\) Veraset provided sufficient pre\- and post\-event mobility coverage, and \(iii\) interference from concurrent or consecutive disasters could be minimized\. The reference event datet0t\_\{0\}, study area, and hazard severity of each event were determined from official hazard\-specific reports and classifications\. Table[3](https://arxiv.org/html/2611.22133#A2.T3)summarizes each event’s metadata, analysis window, and data sources\. All future weather come from a single forecast issued at the forecast origin \(t0\+5t\_\{0\}\+5\)\. Beyond a product’s forecast range, we keep its last value unchanged\. For example, NWM flood forecasts cover 240 h, so in the 14\-day test, days 11–14 reuse the day\-10 value\.

#### B\.1\.2\.Mobility Processing\.

For each event, raw visit records were mapped to U\.S\. Census tracts and ordered by device and timestamp\. Consecutive visits to different tracts were converted into directed movements and aggregated by day to obtain the daily inter\-tract flowsfi​j,e​\(τ\)f\_\{ij,e\}\(\\tau\)\. Following Definition 2\.1, the total mobility of each tract was calculated by summing its daily incoming and outgoing inter\-tract movements\. We then estimated tract\-specific mobility baselines using the two complete weeks preceding the observation window,\[τ0e−23,τ0e−10\]\[\\tau\_\{0\}^\{e\}\-23,\\tau\_\{0\}^\{e\}\-10\]\. Separate weekday and weekend baselines were used to account for calendar effects\. Tracts with fewer than 50 mean daily inter\-tract movements during this baseline period were excluded to reduce sampling bias in sparsely observed areas\. The remaining daily mobility values were normalized by their corresponding weekday or weekend baselines and subsequently processed using a three\-day causal moving average\.

#### B\.1\.3\.Temporal Window Design\.

For each event, data were aligned to the reference datet0t\_\{0\}and collected fromt0−23t\_\{0\}\-23tot0\+19t\_\{0\}\+19\. The interval\[t0−23,t0−10\]\[t\_\{0\}\-23,t\_\{0\}\-10\]forms the 14\-day baseline period,\[t0−9,t0\+5\]\[t\_\{0\}\-9,t\_\{0\}\+5\]the 15\-day observation window, and\[t0\+6,t0\+19\]\[t\_\{0\}\+6,t\_\{0\}\+19\]the maximum prediction horizon used in sensitivity analysis\. The primary seven\-day horizon spans\[t0\+6,t0\+12\]\[t\_\{0\}\+6,t\_\{0\}\+12\]\. The observation window captures the maximum impact and early recovery signals used to construct the router features\([Hsu and Mostafavi, 2025](https://arxiv.org/html/2611.22133#bib.bib22)\)\. Early\-recovery is defined at the tract level as the first rebound of at least10%10\\%of the maximum impact depth toward baseline\. A post\-hoc check confirmed that over95%95\\%of the samples met the required criteria\. The seven\-day primary prediction horizon is motivated by two considerations: reported recovery intervals generally span 18–168 hours\([Donovan and Work, 2017](https://arxiv.org/html/2611.22133#bib.bib12)\), while horizons over a week extend beyond the reliable weather forecast window and may introduce additional noise into the model\.

Table 3\.Event inventory, common data\-collection windows, and hazard\-data sources for the 18 disaster events\.EventAffected
tractsEvent start
date \(t0t\_\{0\}\)Data collection
time frameAffected
usersDisaster
intensityData sources and rolesHurricanesDorian \(2019\)242019/9/62019/8/14–2019/9/2560,341RainfallHurricane reports: NHC\([NOAA National Hurricane Center, nd](https://arxiv.org/html/2611.22133#bib.bib50)\)Hurricane tracks: HURDAT2\([NOAA National Hurricane Center, 2026](https://arxiv.org/html/2611.22133#bib.bib49)\)Official affected areas: FEMA\([Federal Emergency Management Agency, nd](https://arxiv.org/html/2611.22133#bib.bib14)\)Past weather: NOAA PSL\([NOAA Physical Sciences Laboratory, nd](https://arxiv.org/html/2611.22133#bib.bib51)\)Forecast weather: GFS\([NOAA National Centers for Environmental Information, nda](https://arxiv.org/html/2611.22133#bib.bib47)\)Laura \(2020\)4642020/8/272020/8/4–2020/9/15510,594Ida \(2021\)4032021/8/292021/8/6–2021/9/17319,843Nicholas \(2021\)1272021/9/142021/8/22–2021/10/3218,065Ian \(2022\)3412022/9/282022/9/5–2022/10/17376,873Francine \(2024\)3982024/9/112024/8/19–2024/9/301,171,640Helene \(2024\)2112024/9/272024/9/4–2024/10/16635,960Milton \(2024\)4912024/10/92024/9/16–2024/10/281,929,541Winter stormsUri \(2021\)7862021/2/142021/1/22–2021/3/51,080,699Temperature dropWinter\-storm reports: NWS\([National Weather Service, nd](https://arxiv.org/html/2611.22133#bib.bib45)\)Official affected areas: FEMA\([Federal Emergency Management Agency, nd](https://arxiv.org/html/2611.22133#bib.bib14)\)Past weather: GHCN\([NOAA National Centers for Environmental Information, ndb](https://arxiv.org/html/2611.22133#bib.bib48)\)Forecast weather: GFS\([NOAA National Centers for Environmental Information, nda](https://arxiv.org/html/2611.22133#bib.bib47)\)Buffalo lake\-effect snowstorm \(2022\)1852022/11/182022/10/26–2022/12/7271,682January winter storm \(2024\)3742024/1/152023/12/23–2024/2/32,705,099January winter storm \(2025\)5432025/1/212024/12/29–2025/2/92,068,469FloodsKY–MO riverine flood \(2022\)3912022/7/282022/7/5–2022/8/16622,394StreamflowFlood reports: NWS\([National Weather Service, nd](https://arxiv.org/html/2611.22133#bib.bib45)\)Official affected areas: FEMA\([Federal Emergency Management Agency, nd](https://arxiv.org/html/2611.22133#bib.bib14)\)Past weather: USGS NWIS\([U\.S\. Geological Survey, nd](https://arxiv.org/html/2611.22133#bib.bib68)\)Forecast weather: NWM\([National Oceanic and Atmospheric Administration, ndb](https://arxiv.org/html/2611.22133#bib.bib44)\)Vermont–New York flood \(2023\)1082023/7/102023/6/17–2023/7/2940,419WildfiresSaddleridge \(2019\)272019/10/102019/9/17–2019/10/29119,149Near\-Surface SmokeWildfire reports: NIFC\([National Interagency Fire Center, nd](https://arxiv.org/html/2611.22133#bib.bib41)\)Fire perimeters: NIFC\([National Interagency Fire Center, nd](https://arxiv.org/html/2611.22133#bib.bib41)\)Past weather: HRRR\-Smoke\([National Oceanic and Atmospheric Administration, nda](https://arxiv.org/html/2611.22133#bib.bib43)\)Forecast weather: HRRR\-Smoke\([National Oceanic and Atmospheric Administration, nda](https://arxiv.org/html/2611.22133#bib.bib43)\)August Complex \(2020\)882020/8/162020/7/24–2020/9/472,919Lahaina \(2023\)452023/8/82023/7/16–2023/8/2729,646Palisades–Eaton \(2025\)252025/1/72024/12/15–2025/1/26117,850
Note\.Because the datasets vary by event, only the source platforms are listed here\. Detailed event\-level sources are available in our repository\.

### B\.2\.Post\-hoc Recovery Type Identification

#### B\.2\.1\.Clustering the recovery curves\.

Unlike studies that cluster complete recovery processes\([Hsu and Mostafavi, 2025](https://arxiv.org/html/2611.22133#bib.bib22);[Hong et al\., 2021b](https://arxiv.org/html/2611.22133#bib.bib20)\), we use a fixed 22\-day window,\[τ0e−9,τ0e\+12\]\[\\tau\_\{0\}^\{e\}\-9,\\tau\_\{0\}^\{e\}\+12\], to ensure consistent temporal coverage across events\. Using the mobility curves processed as described above, the clustering input for each tract–event sample is

\(25\)𝐱i,e=\[m~i,esm​\(τ0e−9\),…,m~i,esm​\(τ0e\+12\)\]⊤∈ℝ22\.\\mathbf\{x\}\_\{i,e\}=\\left\[\\widetilde\{m\}^\{\\mathrm\{sm\}\}\_\{i,e\}\(\\tau\_\{0\}^\{e\}\-9\),\\ldots,\\widetilde\{m\}^\{\\mathrm\{sm\}\}\_\{i,e\}\(\\tau\_\{0\}^\{e\}\+12\)\\right\]^\{\\top\}\\in\\mathbb\{R\}^\{22\}\.We use DTW K\-means because modest differences in disruption and recovery timing may cause otherwise similar curves to be misclassified under pointwise Euclidean distance\. For any two clustering inputs𝐱\\mathbf\{x\}and𝐲\\mathbf\{y\}, their DTW distance is

\(26\)dDTW​\(𝐱,𝐲\)=minπ⁡\[∑\(u,v\)∈π\(xu−yv\)2\]1/2,d\_\{\\mathrm\{DTW\}\}\(\\mathbf\{x\},\\mathbf\{y\}\)=\\min\_\{\\pi\}\\left\[\\sum\_\{\(u,v\)\\in\\pi\}\(x\_\{u\}\-y\_\{v\}\)^\{2\}\\right\]^\{1/2\},whereuuandvvindex days in𝐱\\mathbf\{x\}and𝐲\\mathbf\{y\}, respectively, andπ\\piis a valid warping path constrained by\|u−v\|≤2\|u\-v\|\\leq 2\. This allows small timing differences without aligning physically distinct recovery dynamics\. For candidate cluster numberKK, the within\-cluster distortion is

\(27\)DK=∑k=1K∑𝐱i,e∈CkdDTW​\(𝐱i,e,𝝁k\)2,D\_\{K\}=\\sum\_\{k=1\}^\{K\}\\sum\_\{\\mathbf\{x\}\_\{i,e\}\\in C\_\{k\}\}d\_\{\\mathrm\{DTW\}\}\\left\(\\mathbf\{x\}\_\{i,e\},\\boldsymbol\{\\mu\}\_\{k\}\\right\)^\{2\},whereCkC\_\{k\}and𝝁k\\boldsymbol\{\\mu\}\_\{k\}are clusterkkand its DTW barycenter\. We use K\-means\+\+ initialization and multiple random restarts\. Then, we evaluateK=2,…,8K=2,\\ldots,8using the DTW\-based silhouette coefficient

\(28\)s⁡\(i\)=b⁡\(i\)−a⁡\(i\)max⁡\{a⁡\(i\),b⁡\(i\)\},s\(i\)=\\frac\{b\(i\)\-a\(i\)\}\{\\max\\\{a\(i\),b\(i\)\\\}\},wherea⁡\(i\)a\(i\)is the mean within\-cluster distance andb⁡\(i\)b\(i\)is the minimum mean distance to another cluster\. We selectedK=4K=4considering the average silhouette score, the elbow inDKD\_\{K\}and requiring that each cluster span multiple events rather than be dominated by a single large event\. The clustering is stable\. The silhouette score is highest atK=4K=4\(0\.34 vs\. 0\.23–0\.31\), and the clusters change little across 10 random seeds, 100 event\-stratified bootstrap resamples, and 18 leave\-one\-event\-out runs \(mean adjusted Rand index of 0\.90, 0\.87, and 0\.83, respectively\), with all four recovery types preserved\. The choice ofKKis independent of the number of ReMoE experts\.

#### B\.2\.2\.Clustered recovery types\.

After selectingK=4K=4, we represent each cluster by the pointwise mean of its event\-aligned curves\. As shown in Figure[3](https://arxiv.org/html/2611.22133#acmlabel2)\(a\), the four recovery types are:

- •Type I: Rapid rebound\.A mild mobility disruption followed by a rapid V\-shaped return to baseline, commonly associated with short\-duration hurricanes or winter storms where mobility resumes quickly after the hazard passes\.
- •Type II: Long deficit\.A pronounced disruption followed by persistently low mobility and limited recovery, typically observed in severely affected tracts following destructive floods, or wildfires where infrastructure damage, service interruption, or displacement is prolonged\.\([Tang et al\., 2023](https://arxiv.org/html/2611.22133#bib.bib66)\)\.
- •Type III: Staged recovery\.An initial rapid rebound followed by a clear reduction in recovery rate\. This pattern often arises after hurricanes or floods when emergency access is restored quickly but broader infrastructure, businesses, and services reopen in stages\([Hong et al\., 2021a](https://arxiv.org/html/2611.22133#bib.bib19)\)\.
- •Type IV: Slow monotonic recovery\.A slow, approximately monotonic return toward baseline with a low initial rebound, typically observed in severely affected tracts with substantial functional loss\.\([Hsu and Mostafavi, 2025](https://arxiv.org/html/2611.22133#bib.bib22)\)\.

## Appendix CRouter Feature Construction

Table 4\.Definitions and sources of the 16 router features used by ReMoE\.CategoryTypeMetricFormula/SourceDefinitionImpactMagnitudeDIDIi,e=\[1−𝐦i,ein,min\]\+\\displaystyle\\mathrm\{DI\}\_\{i,e\}=\\left\[1\-\{\\mathbf\{m\}\}^\{\\mathrm\{in\},\\min\}\_\{i,e\}\\right\]\_\{\+\}Maximum post\-event mobility loss within the early observation window\.DurationTimpTimpi,e=τi,emin−τ0eτine−τ0e\\displaystyle\\mathrm\{Timp\}\_\{i,e\}=\\frac\{\\tau^\{\\min\}\_\{i,e\}\-\\tau\_\{0\}^\{e\}\}\{\\tau\_\{\\mathrm\{in\}\}^\{e\}\-\\tau\_\{0\}^\{e\}\}Normalized time required to reach the maximum observed impact\.IntegralImpPersistImpPersisti,e=∫τ0eτi,emin\[1−𝐦i,ein​\(τ\)\]\+​𝑑τDIi,e​\(τi,emin−τ0e\)\\displaystyle\\mathrm\{ImpPersist\}\_\{i,e\}=\\frac\{\\int\_\{\\tau\_\{0\}^\{e\}\}^\{\\tau^\{\\min\}\_\{i,e\}\}\[1\-\{\\mathbf\{m\}\}^\{\\mathrm\{in\}\}\_\{i,e\}\(\\tau\)\]\_\{\+\}\\,d\\tau\}\{\\mathrm\{DI\}\_\{i,e\}\(\\tau^\{\\min\}\_\{i,e\}\-\\tau\_\{0\}^\{e\}\)\}Normalized cumulative impact before the observed minimum; larger values indicate a more persistent impact phase\.RateRelFRRelFRi,e=\[𝐦i,ein​\(τ0e\)−𝐦i,ein,min\]\+max⁡\(DIi,e,ϵ\)​max⁡\(Timpi,e,0\.2\)\\displaystyle\\mathrm\{RelFR\}\_\{i,e\}=\\frac\{\[\{\\mathbf\{m\}\}^\{\\mathrm\{in\}\}\_\{i,e\}\(\\tau\_\{0\}^\{e\}\)\-\{\\mathbf\{m\}\}^\{\\mathrm\{in\},\\min\}\_\{i,e\}\]\_\{\+\}\}\{\\max\(\\mathrm\{DI\}\_\{i,e\},\\epsilon\)\\max\(\\mathrm\{Timp\}\_\{i,e\},0\.2\)\}Relative speed at which the observed mobility loss is realized\.CensoringCensoredCensoredi,e=𝕀⁡\(τi,emin=τine\)\\displaystyle\\mathrm\{Censored\}\_\{i,e\}=\\mathbb\{I\}\\\!\\left\(\\tau^\{\\min\}\_\{i,e\}=\\tau\_\{\\mathrm\{in\}\}^\{e\}\\right\)Indicates that the observed minimum occurs at the observation boundary\.RecoveryRateRRRRi,e=𝐦i,ein​\(τine\)−𝐦i,ein,minmax⁡\(τine−τi,emin,1\)\\displaystyle\\mathrm\{RR\}\_\{i,e\}=\\frac\{\{\\mathbf\{m\}\}^\{\\mathrm\{in\}\}\_\{i,e\}\(\\tau\_\{\\mathrm\{in\}\}^\{e\}\)\-\{\\mathbf\{m\}\}^\{\\mathrm\{in\},\\min\}\_\{i,e\}\}\{\\max\(\\tau\_\{\\mathrm\{in\}\}^\{e\}\-\\tau^\{\\min\}\_\{i,e\},1\)\}Mean recovery rate from the observed minimum to the observation endpoint\.Momentum𝐦i,ein​\(τ0e\+d\)=αi,e\+βi,e​d\+εi,e,d,d∈\{2,3,4,5\};Momentumi,e=β^i,e\\displaystyle\{\\mathbf\{m\}\}^\{\\mathrm\{in\}\}\_\{i,e\}\(\\tau\_\{0\}^\{e\}\+d\)=\\alpha\_\{i,e\}\+\\beta\_\{i,e\}d\+\\varepsilon\_\{i,e,d\},\\hskip 8\.50012ptd\\in\\\{2,3,4,5\\\};\\hskip 8\.50012pt\\mathrm\{Momentum\}\_\{i,e\}=\\widehat\{\\beta\}\_\{i,e\}Recent recovery slope near the observation endpoint\.RatioRRatioRRatioi,e=𝐦i,ein​\(τine\)−𝐦i,ein,minmax⁡\(DIi,e,ϵ\)\\displaystyle\\mathrm\{RRatio\}\_\{i,e\}=\\frac\{\{\\mathbf\{m\}\}^\{\\mathrm\{in\}\}\_\{i,e\}\(\\tau\_\{\\mathrm\{in\}\}^\{e\}\)\-\{\\mathbf\{m\}\}^\{\\mathrm\{in\},\\min\}\_\{i,e\}\}\{\\max\(\\mathrm\{DI\}\_\{i,e\},\\epsilon\)\}Fraction of the maximum observed mobility loss restored by the observation endpoint\.IntegralRecDefRatioRecDefRatioi,e=∫τi,eminτine\[1−𝐦i,ein​\(τ\)\]\+​𝑑τmax⁡\(DIi,e,ϵ\)​max⁡\(τine−τi,emin,1\)\\displaystyle\\mathrm\{RecDefRatio\}\_\{i,e\}=\\frac\{\\int\_\{\\tau^\{\\min\}\_\{i,e\}\}^\{\\tau\_\{\\mathrm\{in\}\}^\{e\}\}\[1\-\{\\mathbf\{m\}\}^\{\\mathrm\{in\}\}\_\{i,e\}\(\\tau\)\]\_\{\+\}\\,d\\tau\}\{\\max\(\\mathrm\{DI\}\_\{i,e\},\\epsilon\)\\max\(\\tau\_\{\\mathrm\{in\}\}^\{e\}\-\\tau^\{\\min\}\_\{i,e\},1\)\}Remaining below\-baseline recovery burden normalized by impact depth and post\-minimum duration\.ABABi,e=∫τi,eminτine\[𝐦i,ein​\(τ\)−1\]\+​𝑑τ\\displaystyle\\mathrm\{AB\}\_\{i,e\}=\\int\_\{\\tau^\{\\min\}\_\{i,e\}\}^\{\\tau\_\{\\mathrm\{in\}\}^\{e\}\}\[\{\\mathbf\{m\}\}^\{\\mathrm\{in\}\}\_\{i,e\}\(\\tau\)\-1\]\_\{\+\}\\,d\\tauCumulative above\-baseline mobility during early recovery\.StabilityPreVarPreVari,e=SD⁡\{𝐦i,ein​\(τ0e−9\),…,𝐦i,ein​\(τ0e−1\)\}\\displaystyle\\mathrm\{PreVar\}\_\{i,e\}=\\operatorname\{SD\}\\\!\\left\\\{\{\\mathbf\{m\}\}^\{\\mathrm\{in\}\}\_\{i,e\}\(\\tau\_\{0\}^\{e\}\-9\),\\ldots,\{\\mathbf\{m\}\}^\{\\mathrm\{in\}\}\_\{i,e\}\(\\tau\_\{0\}^\{e\}\-1\)\\right\\\}Pre\-event variability of normalized mobility\.ContextualPressureSTPobs\\mathrm\{STP\}^\{\\mathrm\{obs\}\}NOAA\([National Oceanic and Atmospheric Administration, \[n\. d\.\]](https://arxiv.org/html/2611.22133#bib.bib42)\)\(see Eq\. \([5](https://arxiv.org/html/2611.22133#S2.E5)\) for details\)Mean observed spatiotemporal pressure from the mobility minimum to the observation endpoint\.VulnerabilitySVICDC/ATSDR\([CDC/ATSDR, 2018](https://arxiv.org/html/2611.22133#bib.bib7)\)Tract\-level social vulnerability\.ExposurePopDensUSEPA\([Chapman et al\., 2021](https://arxiv.org/html/2611.22133#bib.bib8)\)Log\-transformed tract population density\.Land useBuiltUpUSEPA\([Chapman et al\., 2021](https://arxiv.org/html/2611.22133#bib.bib8)\)Share of developed or built\-up land within the tract\.Network centralityPRbaseEq\. \([30](https://arxiv.org/html/2611.22133#A3.E30)\)Event\-level percentile of baseline weighted PageRank\.
Notation\.We defineτi,emin\\tau^\{\\min\}\_\{i,e\}as the earliest day whenm~i,ein\\widetilde\{m\}^\{\\mathrm\{in\}\}\_\{i,e\}reaches its minimum over\[τ0e,τine\]\[\\tau\_\{0\}^\{e\},\\tau\_\{\\mathrm\{in\}\}^\{e\}\], denote the corresponding value bym~i,ein,min\\widetilde\{m\}^\{\\mathrm\{in\},\\min\}\_\{i,e\}, and use\[x\]\+=max⁡\(x,0\)\[x\]\_\{\+\}=\\max\(x,0\)withϵ=10−4\\epsilon=10^\{\-4\}\.

Table[4](https://arxiv.org/html/2611.22133#A3.T4)details the 16 router features\. Impact and recovery features are derived from the predefined 15\-day observation window\[τ0e−9,τ0e\+5\]\[\\tau\_\{0\}^\{e\}\-9,\\tau\_\{0\}^\{e\}\+5\]\. To avoid redundancy, we retain impact depth but exclude residual mobility at maximum impact, as one is mathematically determined by the other\. Similarly, normalized impact timing and relative decline rate are used instead of their absolute counterparts to improve cross\-event comparability\. For contextual features, we average the daily STP scores over the observation window\. Region\-related features, including social vulnerability, population density, and built\-environment characteristics, are obtained from the datasets listed in Table[4](https://arxiv.org/html/2611.22133#A3.T4)\. Additionally, we include baseline network centrality to capture differences in regions’ pre\-event network positions\. Specifically, using the mobility flows over the two weeks preceding the observation window, we construct

\(29\)wj​i,ebase=114​∑τ=τ0e−23τ0e−10fj​i,e​\(τ\),j≠i\.w^\{\\mathrm\{base\}\}\_\{ji,e\}=\\frac\{1\}\{14\}\\sum\_\{\\tau=\\tau\_\{0\}^\{e\}\-23\}^\{\\tau\_\{0\}^\{e\}\-10\}f\_\{ji,e\}\(\\tau\),\\qquad j\\neq i\.The baseline PageRank percentile is then

\(30\)PRbasei,e=PctRanke⁡\[PageRank0\.85⁡\(𝐖ebase\)i\]\.\\mathrm\{PRbase\}\_\{i,e\}=\\operatorname\{PctRank\}\_\{e\}\\left\[\\operatorname\{PageRank\}\_\{0\.85\}\\left\(\\mathbf\{W\}^\{\\mathrm\{base\}\}\_\{e\}\\right\)\_\{i\}\\right\]\.A higherPRbasei,e\\mathrm\{PRbase\}\_\{i,e\}indicates a more central region\.

Finally, we assessed feature redundancy using Rank\-VIF, which captures redundancy arising from monotonic relationships\. As shown in Figure[6](https://arxiv.org/html/2611.22133#acmlabel5), most Rank\-VIF values were below 10; RRatio was retained despite a slight exceedance because it captures distinct recovery progress\.

![Rank-based variance inflation factors for the 16 router features, grouped into impact, recovery, and contextual categories.](https://arxiv.org/html/2611.22133v1/figures/feature_filtering_diagnostics.png)Figure 6\.Rank\-VIF diagnostics for the 16 router features\.Rank\-based variance inflation factors for the 16 router features, grouped into impact, recovery, and contextual categories\.
## Appendix DAdditional Experiments

- •EQ\-A1:How do different feature groups affect ReMoE’s forecasting performance?→\\rightarrowSec\.[D\.1](https://arxiv.org/html/2611.22133#A4.SS1)
- •EQ\-A2:Does ReMoE learn a routing space better aligned with recovery semantics than a standard MoE?→\\rightarrowSec\.[D\.2](https://arxiv.org/html/2611.22133#A4.SS2)
- •EQ\-A3:How do input features contribute nonlinearly to ReMoE’s expert assignment?→\\rightarrowSec\.[D\.3](https://arxiv.org/html/2611.22133#A4.SS3)

### D\.1\.Impact of Feature Group on Model Performance \(EQ\-A1\)

![Violin plots showing tract-level RMSE distributions for the full model and models without impact, recovery, or contextual router features.](https://arxiv.org/html/2611.22133v1/figures/appendix_additonal_01.png)Figure 7\.Router feature\-group ablation\.Tract\-level RMSE distributions after neutralizing each router feature group\.Violin plots showing tract\-level RMSE distributions for the full model and models without impact, recovery, or contextual router features\.Here we present a feature\-level supplementary ablation to evaluate the value of features groups used by router\. “Full” denotes the complete model using all 16 router features, while “w/o Impact,” “w/o Recovery,” and “w/o Contextual” denote variants with the corresponding impact, recovery, or contextual feature group removed, with all other inputs unchanged\.

Figure[7](https://arxiv.org/html/2611.22133#acmlabel6)shows the RMSE distributions on the ID test set, where red lines indicate the mean and black lines indicate the mean plus one standard deviation\. The full model achieves a mean RMSE of 0\.055\. Removing impact features increases the mean RMSE to 0\.061 \(\+10\.9%\), removing contextual features to 0\.060 \(\+9\.1%\), and removing recovery features produces the largest degradation, reaching 0\.064 \(\+16\.4%\)\. This ordering suggests that recovery features may carry the largest aggregate attribution, followed by impact and contextual features, which provide complementary but smaller contributions to expert routing and forecasting\. Nevertheless, retaining all feature groups yields the best overall performance\.

### D\.2\.Routing\-Space Comparison \(EQ\-A2\)

![Refer to caption](https://arxiv.org/html/2611.22133v1/figures/appendix_additonal_03.png)Figure 8\.Routing\-space comparison\.t\-SNE visualization of the routing spaces learned by ReMoE \(a\) and a standard MoE \(b\), with samples colored by post hoc recovery type\.We examine whether the integrated MoE design of ReMoE learns a more semantically structured routing space than a standard MoE\. The standard MoE uses a latent\-representation\-based router and conventional MLP experts\. For both models, we project the pre\-Top\-KKrouter logits using t\-SNE and color each sample by its post hoc recovery type\. As visualized in Figure[8](https://arxiv.org/html/2611.22133#A4.F8), ReMoE forms more coherent regions aligned with different recovery types, whereas the standard MoE exhibits greater mixing\. This suggests that ReMoE not only makes expert selection easier to trace, but also yields a routing space more closely aligned with post hoc recovery types\.

We quantify this effect using the Calinski–Harabasz \(CH\) index, which measures cluster compactness and separation\. CH is computed on the original five\-dimensional router logits for both models\. Higher CH values indicate better\-defined clusters\. ReMoE achieves a substantially higher CH score than the standard MoE \(735\.98 vs\. 30\.47\), supporting a more structured routing space\.

### D\.3\.Nonlinear Feature Effects on Expert Routing \(EQ\-A3\)

![Refer to caption](https://arxiv.org/html/2611.22133v1/figures/appendix_additonal_02.png)Figure 9\.Global router attribution of the 16 input features\.To complement the monotonic associations in Fig\.[3](https://arxiv.org/html/2611.22133#acmlabel2)d, we apply GradientSHAP to quantify the nonlinear contributions of the 16 input features to each sample’s dominant expert, using its pre\-Top\-KKsoftmax probability as the target\. The left beeswarm in Figure[9](https://arxiv.org/html/2611.22133#A4.F9)shows signed attributions to each sample’s dominant expert, with color indicating feature magnitude\. The horizontal axis measures how each feature changes the dominant expert’s routing weight\. For example, high Momentum values are concentrated on the positive side, indicating that stronger recovery momentum generally reinforces the dominant\-expert assignment\. In contrast, for most features, including DI, PreVar, ImpPersist, and RecDef, both high and low values appear on either side of zero\. It is understandable that the same feature may strengthen the dominant\-expert weight for some samples but weaken it for others, revealing nonlinear and context\-dependent routing behavior\.

The right panel shows that impact depth has the largest mean absolute attribution \(15\.0%\), followed by momentum \(11\.7%\) and built\-up area \(11\.2%\)\. Pre\-event variability, impact persistence, and recovery deficit each contribute approximately 9%, while recovery rate, observed STP, population density, and the baseline restored ratio contribute 7\.9%, 6\.2%, 5\.8%, and 4\.2%, respectively\. At the feature\-group level, recovery, contextual, and impactfeatures account for 38\.4%, 31\.4%, and 30\.2% of the total attribution\. Thus, all three information groups play important roles in expert assignment,with recovery features contributing slightly more overall\.

For implementation, we use Captum’s GradientSHAP\([Kokhlikyan et al\., 2020](https://arxiv.org/html/2611.22133#bib.bib32)\)with an event\-balanced background of 200 training samples, 1,024 Monte Carlo samples, and Gaussian noise \(σ=0\.01\\sigma=0\.01\)\. 95% confidence intervals are obtained from 5,000 event\-level bootstrap replicates\.

Similar Articles

Distributionally Robust Mixture-of-Experts Training

Hugging Face Daily Papers

提出DRMoET——一种将各层专家视为内生鲁棒性分组、优化高损失路由结果的即插即用MoE训练目标,在10.3B总参数规模上显著提升下游任务平均表现与路由鲁棒性。

E-MoE: Enhanced Mixture-of-Experts for Non-Factorized Diffusion Language Models

Hugging Face Daily Papers

论文提出 E-MoE(Enhanced Mixture-of-Experts),用 MoE 路由决策构成的离散共享隐变量来建模非因子化的反向过程,缓解连续隐变量的后验坍缩问题,并在无需增加活跃参数的情况下提升掩码扩散语言模型在少步生成场景下的样本质量。