Unveiling multifaceted resilience: A heterogeneous graph neural network approach for analyzing locale recovery patterns

Abstract

Resilience, denoting the capacity to swiftly recover to a state of normalcy subsequent to the occurrence of a disaster, constitutes a multifaceted phenomenon necessitating in-depth investigation. This study undertakes the quantification of resilience pertaining to specific locales through the utilization of heterogeneous data encompassing visitation patterns, demographic particulars, and points of interest (POI). A heterogeneous graph neural network is applied to model the resilience of these locales in Galveston, TX, USA. Our model outperforms regression models and other homogeneous baseline methodologies. Subsequent analysis unveils discernible resilience patterns intertwined with metrics such as visitation frequencies, visitors’ travel behaviors, and geographical attributes. In comparison to resilience investigations solely predicated upon visitation counts, our approach captures a more extensive array of information, thereby yielding a comprehensive understanding of the locale’s resilience.

Keywords

Resilience point of interest graph neural networks deep learning GeoAI

Introduction

Resilience, defined as a community’s ability to endure and rebound from disruptions, holds immense importance in urban planning, public policy, and disaster management (Adger, 2000; Cutter et al., 2008; Holling, 1973; Ye et al., 2023). Traditional approaches to assessing resilience often rely on static indicators like population density, land use, etc (Amin et al., 2022; Cai et al., 2018). Yet, these indicators might not fully capture resilience’s intricate and dynamic nature.

Recently, integrating human mobility data into resilience studies has gained traction (Barbosa et al., 2018; Heydari et al., 2023; Hu et al., 2021; Huang et al., 2021; Yabe et al., 2020). This dimension of resilience recognizes people’s adaptive capacities and choices in adversity. An example of such data is Point of Interest (POI) visitation data, which records foot traffic at specific places like businesses, parks, and cultural sites, reflecting local activity and vibrancy (Song et al., 2022; Zhao et al., 2023). POI visitation data offers insights into recovery by revealing the speed and extent of people returning to normal activities. For instance, after a natural disaster, a drop in POI visits suggests reduced resilience. Conversely, a swift return to pre-disruption levels indicates strong resilience. This data also reflects community well-being; higher POI visits imply confidence and safety. Conversely, limited visits may signal insecurity, impacting community resilience (Podesta et al., 2021; Wang et al., 2023).

Studying recovery patterns is crucial for policymakers to make informed decisions to enhance resilience (Berke et al., 2015). Current deep learning methods have demonstrated their potential to enhance our grasp of resilience in intricate urban environments. They’ve been applied to predict resilience-related outcomes such as recovery patterns post-disaster, infrastructure vulnerability, and community adaptability (Shin et al., 2017; Yuan et al., 2022; Zou et al., 2018). Despite their strengths, these approaches might not be optimally tailored to address the distinct challenges of human mobility and resilience. Specifically, they could struggle with capturing the inherent spatial and temporal interdependencies crucial to understanding people’s movements and adaptations during disruptions.

This is where Graph Neural Networks (GNNs) offer a promising alternative. GNNs are a specialized category of machine learning models designed to dissect and extract insights from intricate interconnected data structures. These structures, often represented as graphs, intricately capture the spatial relationships and interdependencies among different urban locations (Zhu et al., 2020). By harnessing GNNs, we can effectively model the intricate interplay between human mobility and resilience, revealing insights that may elude traditional machine learning or deep learning methods.

In this paper, we aim to characterize POIs by leveraging comprehensive mobility data and place attributes in a natural disaster. We propose a graph model utilizing heterogeneous GNNs to analyze and compute resilience patterns. Our POI-human mobility graph is designed as a versatile and scalable model that can incorporate interconnected urban data, allowing for possible continuous enrichment with new information. This approach uncovers unique spatial resilience patterns and insights, complementing prior mobility-focused resilience studies with advanced methods. This new method may enhance discussions on resilience and guide strategies for community preparedness against increasing disaster and climate change challenges.

Our contributions are threefold: (1) Conceptually, we analyze resilience within a graph-based framework that harmonizes diverse datasets, encompassing economic, demographic, and human mobility data. This innovative approach provides a holistic method for evaluating place resilience. (2) Technically, we employ a heterogeneous graph neural network to model the POI graph. This deep learning model adeptly captures non-linear relationships in our heterogeneous urban datasets, furnishing a robust tool for resilience analysis. (3) Empirically, we apply our methodology to Galveston, TX, during Winter Storm Uri. This experiment decomposes the resilience curve and uncovers two distinct patterns of place resilience.

In subsequent sections, we delve into our literature review, detail our utilized data, introduce our graph-based deep learning model for resilience, present our experimental results and the Galveston, TX case study during Winter Storm Uri, and conclude our study.

Literature review

Business resilience, natural hazards, and human mobility

Business resilience refers to the ability of businesses to absorb, recover, and adapt to adverse conditions, such as natural hazards (Adger, 2000; Cutter et al., 2008; Holling, 1973). Saad et al. (2021) provided a comprehensive review on business resilience research, covering the varying definitions and measurements. Resilience has gained increased research attention recently, with a strong focus on its integration with sustainability and development (Mehta et al., 2024).

Business resilience encompasses multiple complex factors (Berke et al., 2015; Cai et al., 2018). Notably, traffic flows, demographic characteristics, and place-specific information emerge as three related indicators for resilience (John Van de Lindt et al., 2018; Wang et al., 2022). For example, Qiang and Xu (2020) used crowdsourced traffic data to assess resilience, defined based on accumulated accessibility reduction over time during a hazard, that is, the quicker normal traffic return, the more place resilience would be. The results reveal strong spatial variations of the resilience in Cleveland, Ohio, USA, during the winter storm Harper. Despite the effectiveness of this resilience assessment method in quantitatively measuring resilience, it omits the internal difference in POIs and fails to consider other environmental factors in the measurement method. Rosenheim et al. (2021) developed the household-level resilience measurement, but their model missed the human mobility data we discussed. Those studies, due to their concepts and methodology limits, put challenges to scale and adopt to study the resilience for new places.

Various approaches exist to link different datasets to resilience, such as direct comparisons of attributes across different places. Initially, resilience can be described using traffic data or other indicators, and subsequently, researchers can explore differences in resilience between locations by analyzing census data or conducting surveys (Du et al., 2020; Qiang et al., 2020). Haraguchi et al. (2022) reviewed the use of human mobility data in the urban resilience analysis. Linear regression is commonly employed to determine the contribution of each attribute to resilience, while spatial calculations are used to address spatial heterogeneity, enabling the integration of information across different scales (Cai et al., 2018; Malczewski and Jankowski, 2020). More recently, deep learning methods have been proposed to encode location information into embeddings, facilitating the integration of multiple datasets for other research topics (Mai et al., 2020).

Deep learning-based resilience measures

There have been multiple studies on applying deep learning in resilience measure. Pioneers came from using deep learning and remote sensing and other spatial temporal big data to predict the hazards occurrence, then link that to the resilience concepts (Dubey et al., 2022; Haggag et al., 2021; Tan et al., 2021; Ti et al., 2022). These machine learning techniques bring advantages like managing vast, complex datasets and grasping nonlinear correlations between variables (Falco, 2019; Shin et al., 2017). Hao and Wang (2023) uses a spatio-temporal deep learning model, specifically coupling graph convolutional neural network (GCN) and long short-term memory (LSTM), to understand the influence of communities’ social and physical properties on extreme event impacts, demonstrating the potential of data-driven approaches for understanding complex, dynamic, and place-specific community resilience.

Graph structures are used to study resilience because they can link data together for a complex object (Wang and Ye, 2018; Zou et al., 2018). A graph is defined by nodes and edges, in which the nodes represent the study objects and edges show the linkages. The graph neural networks are based on deep learning methods and would be able to learn complex patterns from the nodes and edges. In the real world, the graph may present features such as heterogenous and dynamics, consisting of multiple types of nodes. The heterogeneous graph neural networks are designed to take the heterogeneous nodes into account. Because they are specifically designed for modelling heterogeneous nodes, they usually perform better in various tasks than homogenous networks (Hu et al., 2020). For example, GNNs could be used to analyze mobility data, such as the number and types of trips people take, to understand how people move around and connect with different places. GNNs could also be used to analyze social media data, such as posts and hashtags, to understand how people communicate and share information about their experiences and needs.

However, it is important to acknowledge that current deep learning methods and graph neural networks are not inherently designed for studying place resilience. Our work specifically focuses on place encoding and extends these methods to resilience research. Through this endeavor, our paper aims to bridge the existing gap between place resilience research and graph neural network applications. By incorporating place encoding into graph neural networks, we seek to enhance their capabilities in analyzing and understanding the resilience of specific locations. This extension will enable us to effectively harness the power of graph neural networks to tackle the complexities inherent in resilience studies, particularly when considering the unique characteristics and dynamics of different places.

Our choice of GNNs in this study is driven by three reasons: (1) GNNs adeptly capture the complex spatial dependencies and interactions among various locations, crucial for understanding human mobility and resilience dynamics. (2) GNNs can manage diverse data sources, enabling integration of socio-economic indicators and infrastructure data into our resilience modelling. (3) GNNs excel at modelling intricate, nonlinear relationships between variables, unveiling hidden patterns and interactions critical for comprehending and enhancing place resilience.

Data

We utilized comprehensive data to observe multiple aspects of resilience, focusing on the winter storm Uri. To gain insights into resilience, we collected data from various sources, including traffic data, demographics, and place-specific information.

Winter storm Uri

Winter storm Uri had a significant impact on Texas, a state that rarely experienced such severe winter weather prior to 2021. The infrastructure was ill-prepared for the prolonged winter situation, and a unique aspect of Texas was its separate electric grid managed by the Electric Reliability Council of Texas, which prevented the transfer of power from other grids. Consequently, the electric grid was offline for an extended period, leading to widespread power outages and significant hardships for residents. During the storm, many households lost electricity, and people faced freezing temperatures, resulting in suffering and adverse consequences (Popovich et al., 2021).

Our study period spanned the week in which the winter storm hit and the preceding week (February 3rd to 17th). This judiciously selected timeframe facilitated the demarcation of a control week, representing ambient conditions antecedent to the disturbance. Opting for the immediately preceding week as a reference offers several methodological advantages. Firstly, it ensures a robust comparability between typical (baseline) and atypical (disaster) conditions. Secondly, by confining our analysis to this short duration, we mitigate potential inconsistencies arising from the addition of new mobile devices to the dataset over an extended period. Such additions could introduce variability, complicating the normalization process. Lastly, the confined timeframe aids in minimizing sampling errors associated with visit count normalization, thereby enhancing the precision and reliability of our findings.

Galveston Island was selected as our focus area due to its location in east coastal Texas, which allowed for relatively quicker recovery compared to other regions (Cai et al., 2023; Gao et al., 2023). This decision enables us to study short-term resilience and observe meaningful recovery within our study period. By concentrating on a relatively small number of POIs in Galveston, we can delve into specific place resilience features. Additionally, the coastal region’s dense population and substantial tourism traffic provide diversity in the places we can study, enriching our understanding of resilience in different contexts.

Mobility and supplemental data

To gather human mobility data, we relied on Safegraph (safegraph.com), a company that offers anonymous mobility data derived from sampled GPS trajectories of mobile phones. This data allowed us to understand how many people visited POIs each day and which CBG they came from, giving valuable insights into mobility patterns during the winter storm Uri. The data were also aggregated in this way to protect individuals’ privacy, so we cannot get exact location of their home address.

For demographical and economic data, we sourced information from the American Community Survey 5-year Estimates by Census Block Group (CBG) for the release year 2020. This dataset provided us with recent and relevant data at a granular level, specifically choosing variables such as median household income, age, and race, which are commonly used in resilience studies (Hao and Wang, 2023; Wang et al., 2022).

Node type	Attribute	Encoding detail
POI	Category	Use one-hot-encoding (dummy variables) to represent the POI category
POI	Location	Latitude and longitude of the POI
CBG	Income	Continuous variable representing the median household income of the census block group
	Age	17 integers as the number of people in 5-yesar age groups (0–4, 5–9, …, 80–84, 85+)
	Race	Two integer variables, one representing the count of white individuals and the other representing the count of non-white individuals within the CBG.
	Location	Latitude and longitude of the geometric center of the CBG

Methods

Our approach focuses on integrating diverse data sources to explore resilience. To address the challenge of handling various data types cohesively, we build a graph incorporating POI and census block group (CBG) data. This graph captures visit flows between POIs and CBGs, offering comprehensive insights into the attributes of each POI and its connection to resilience. Leveraging deep learning techniques, we use this graph to generate contextual embeddings representing the strengths of individual POIs and CBGs. Finally, we qualitatively analyze these embeddings to gain a deeper understanding of each POI’s characteristics. Figure 1 shows the workflow of our approach.

Figure 1.

This figure illustrates the overall workflow of our approach to urban resilience modelling. Initially, we constructed a heterogeneous graph that integrates Points of Interest (POIs) and Census Block Groups (CBGs), incorporating both mobility and supplemental data to capture the dynamics of urban interactions. Subsequently, this graph was processed using a heterogeneous graph neural network (HGNN), which was employed to compute embeddings for both CBGs and POIs based on their relational and attribute data. Finally, these embeddings were utilized to predict visitation patterns, identify clusters of POIs, and derive implications for policy-making based on the spatial and functional characteristics of the clusters.

Data preprocessing

Some POI is popular while others are not. The Safegraph sampled user trajectory can reflect this popularity, showing a considerable difference in the visit count. Directly comparing those visit counts may mislead us because those large count values may bias the results. We make the visit count comparable for each POI by only comparing their visit count. For each POI, we use the visit count of Feb 11–17 minus that of Feb 3–10 to calculate the differences in visit count. We implemented a normalization procedure leveraging the min-max scaling technique. In this methodology, the visitation metrics for each POI are scaled based on the extremities of counts observed during the designated research timeframe. Consequently, the resilience trajectory for every individual POI is constrained within the interval [–1, 1]. Within this spectrum, a value of −1 signifies the most pronounced decline in visits ever recorded for that POI, whereas a value of 1 indicates a surge in visitations compared with the previous week. This standardization enables a more accurate and meaningful comparison of resilience measures across various POIs, ensuring that our analysis captures the relative differences in their adaptive capacities.

The heterogeneous POI and CBG graph formalization

The basic units of our data are the POI, the CBG, and the visitation flow between POI and CBG. We can naturally construct a graph with POI and CBG as nodes and the flow as the link for this graph. In this way, the POI-CBG graph describes how humans travel to the POI from their home CBG. We would be able to study the POI resilience not only from the features of the place but also from the visitor’s information.

The CBG contains the demographic information within, so we decided to make the graph attributed so it can model the rich information on the nodes. Similar to the CBG nodes, we add categorical information to the POI nodes. Specifically, each node was represented as a vector $x$ . Each dimension of the vector is an attribute. For the CBG, the attributes are the numerical median income values, population by age, and race. For the POI, the attributes are just the categorical coding for the place, such as restaurants, grocery stores, auto repair shops, etc. We denote the link between the POI and CBG as vector $e$ , containing the number of visits.

Heterogeneous graph neural network

Heterogeneous GNNs are especially fitting for modelling place resilience due to their ability to accommodate various data types and relationships, encompassing individuals, locations, and activities. Through the application of heterogeneous GNNs to resilience modelling, we can capture the intricate and evolving interactions among people, places, and activities that bolster resilience. By amalgamating and scrutinizing these diverse data streams, GNNs can construct intricate and dynamic depictions of place resilience. These depictions, in turn, can inform targeted and well-informed recommendations for enhancing resilience. Examples include identifying critical nodes or pathways in the network vital for resilience, or pinpointing susceptible areas necessitating additional support.

We build our neural network based on the Heterogenous Graph Attention Network (Wang et al., 2019) and implement in the PyG library (Fey and Lenssen, 2019). As Figure 1 shows, the idea is to get the contextual embedding $x'$ for each POI so we can analyze the resilience in a comprehensive way. The high-level idea can be written as

x_{i}^{'} = α_{i, i} Θ x_{i} + \sum_{j \in N (i)} α_{i, j} Θ x_{j}

(1)

{x'}_{i}

is the embedding for node i. The first half part of the information

α_{i, i} Θ x_{i}

is from the node itself; the second half shows the embedding should contain the information from its neighbors

N (i)

Θ

is the learnable parameters in neural network layers (Hamilton et al., 2017). In the graph attention network, learnable parameters

Θ

reside in GATConv, the Graph Attention Convolutional layers, as Figure 1 shows. Those

Θ

are trained different based on their connectivity to POI and CBG, so they are heterogenous and can learn node type specific features. The coefficients

α_{i, i}

are computed as

α_{i, j} = \frac{\exp (a^{⊤} R e L U (Θ [x_{i} ‖ x_{j} ‖ e_{i, j}]))}{\sum_{k \in N (i) \cup i} \exp (a^{⊤} R e L U (Θ [x_{i} ‖ x_{k} ‖ e_{i, k}]))}

(2)

Exp

is the exponential operator.

a^{⊤}

is the transposition of the adjacency matrix in the graph.

L e a k y R e L U

is a non-linear activation function to model the non-linear relationships in the data.

Θ

is MLP to take all the node and edge attributes that also depend on the direction of node types.

‖

is the concatenation operator for vectors. Using the exponential operator in both the nominator and denominator is a standard way to calculate attention. It can also normalize the data and make the equation differentiable, among other admirable characteristics.

We also normalize the visit flow so that the $e$ can show the portion of the visit count instead of the absolute value. This would hide the raw visit count and provide generalizability if the model wants to be deployed for different POI

e_{i, j} = \frac{e_{i, j}}{\sum_{k \in N (i) \cup i} e_{i, k}}

(3)

The above equations (1) and (2) only present one layer of the neural network. By staking another layer, we can aggregate more contextual information. For example, in the first layer, the $x_{i}$ is a CBG and $x_{j}$ is a POI, so the output of first layer is a CBG with the POI information, which shows the travel needs for this CBG. Then in the second layer, $x_{i}$ is a POI and $x_{j}$ is the CBG. The output ${x'}_{i}$ of second layer is a POI embedding that not only contains the information from the POI itself, the demographic and economic information of the coming visitors, and CBG travel demands.

The embedding (contextual representation) ${x'}_{i}$ for each POI were fed into a MLP to predict the resilience curve. We use this to regularize all the learnable parameters in $Θ$ . The objective function is to minimize the Mean Squared Error (MSE) for all the POI

M S E = \frac{1}{n} \sum_{i = 1}^{n} {(Θ x_{i}^{'} - y_{i})}^{2}

(4)

n

is the number of POI,

Θ

is the MLP model,

Θ {x'}_{i}

is the predicted resilience curve,

y_{i}

is the resilience curve of POI i described in Section 4.1. The lower the MSE, the more accurate the model.

Other techniques we use include the dropout, which randomly drop some data in the training process. This dropout mimics the sampling process, as the sampling may miss some POI and visitors. We think the results would be better generalized using the dropout technique.

Embedding calculation and clustering

To compare the embedding of our model with resilience curves utilized in other studies, we introduced two types of distance measures: (1) the distance between differences in visit counts and (2) the distance between embeddings. Since both the visit counts and embeddings are vectors, we leverage the most intuitive Euclidian distance which is $\sqrt{\sum_{i = 1}^{n} {(a_{i} - b_{i})}^{2}}$ , n is the dimension of vector, $a_{i}$ is the ith dimension in two vectors. For the distance between differences in visit counts, the vectors are the visits counts for two POIs. While for the distance between embeddings, the two vectors are the embeddings of the two POIs ${x'}_{i} a n d {x'}_{j}$ , These two distances allowed us to explore the relationships between POIs and their resilience characteristics from two different perspectives. We ran ad hoc analysis to understand the difference between the embedding distance and visitation count.

The embedding of the POI ${x'}_{i}$ can be clustered using t-distributed stochastic neighbor embedding (Van der Maaten and Hinton, 2008). This method is particularly effective in visualizing high-dimensional data by converting similarities between data points into joint probabilities. It aims to place similar embeddings in close proximity while increasing the distance between distinct ones, thus preserving the local structure of the data. Once the embeddings are generated, they can be clustered based on their distances. To accomplish this, we employ the k-means clustering method, a popular unsupervised learning algorithm used for partitioning a dataset into a set of k groups or clusters. The number of clusters, k, was determined using the elbow method, a technique proposed by (Sculley, 2010). The elbow method involves plotting the explained variation as a function of the number of clusters and picking the elbow of the curve as the number of clusters to use. After the clustering process, we delve into the raw data to understand the characteristics of each cluster. This involves examining the properties of the POIs within each cluster, such as their types, locations, and other relevant features. By doing so, we can gain insights into the inherent structure of our data and understand the relationships between different POIs, which is crucial for our subsequent resilience analysis.

Results

We first conducted several experiments to show that our model can achieve better accuracy than baseline models. Then we interpreted the model results using the embeddings of POI. Those analyses would show our model can derive extra knowledge about place resilience compared to traditional methods.

Modelling accuracy

We compared our model with other approaches including linear regression and persistence model. Our model means the heterogeneous graph neural network described in section 4. The homogenous graph neural network converted our graph to a homogenous graph by adding all the information to POI. Besides the categorical information and location coordinates, the POI nodes also contain demographic information from the visitors. Then the homogenous graph neural network learns how that information can contribute to the final prediction. This homogenous graph neural network model did not contain edges but modeled the non-linear relationship between the POI features and the resilience.

On the other hand, the linear regression model uses the same input as the homogenous graph neural network but only uses the linear model to map the relation between the POI features and the resilience. The mean baseline calculates the mean of the resilience curves and compares all the resilience curves with the mean curve. All the models were trained on 70% of the POI and tested on the rest 30% of the POI. We ran five-fold cross-validation and reported the Mean Squared Error (MSE) and Mean Absolute Error (MAE) Table 1.

Table 1.

Comparison of Model Prediction Accuracy (Lower Values Indicate Better Performance; Bolded Values Represent the Best).

	Heterogeneous graph attention network	Homogenous graph neural network	Linear regression	Mean
MSE	0.055	0.094	0.383	0.110
MAE	0.235	0.306	0.619	0.331

The accuracy of our model is best, where the Mean Squared Error (MSE) for the Heterogeneous Graph Attention Network is 0.055, significantly outperforming the Homogenous Graph Neural Network (MSE of 0.094 and MAE of 0.306) and Linear Regression (MSE of 0.383 and MAE of 0.619). This improvement in accuracy underscores the advantages of modelling Points of Interest (POI) and Census Block Groups (CBG) separately in a heterogeneous model. Furthermore, the poor performance of the linear regression model, even when compared to the mean baseline (MSE of 0.110 and MAE of 0.331), indicates the existence of numerous non-linear relationships between the POI features and resilience, which traditional linear models fail to accurately capture.

Clustering

Our model produces the embeddings $x$ for POI, which contains the rich information about the place resilience. The embedding of the POI can be clustered using t-distributed stochastic neighbor embedding (Van der Maaten and Hinton, 2008). We ran the k-means clustering method and got four clusters as the elbow methods suggested (Sculley, 2010). Clusters 0, 1, 2, and 3 are coded red, green, purple, and blue accordingly.

Geographical distribution of the cluster

The clusters showed spatial distribution. As Figure 2 shows, Galveston Island stretches west-east. The north side connects to Houston inland, while the south side is the Gulf of Mexico. The POI clusters 0 and 2 are primarily located in the coastal regions of Galveston. The majority (43 out of 67) of cluster 0 is traveler accommodation. The majority (90 out of 132) of cluster 2 are bars and restaurants. The POI clusters 1 and 3 are located on the island. Although no specific type of POI dominates clusters 1 and 3, there are 20 (out of 121) gas stations in cluster 1 and 29 (out of 125) schools and museums in cluster 3.

Figure 2.

Heterogeneous graph attention network encodes the POI-CBG graph to their corresponding embeddings.

Differences on the resilience curve

We can colorize the previous figure based on the POI cluster schema. As Figure 3 shows, every line is colored according to its cluster. The solid lines are the median for that cluster, while other transparent lines in the background are the raw POI visits differences.

Figure 3.

Spatial distribution of the POI clusters. Node shows the POI locations. Clusters are color-coded.

Two patterns are observed. Cluster 0 and 2 shows one type of visitation change pattern while cluster 1 and 3 shows the other. The first pattern (featuring clusters 0 and (2) of the visit number would decrease first, reaching the lowest on the day of the winter storm (day 3). Then the visit count would recover. The second pattern would keep a stable visit count during the first 2 days of the disaster but decrease the number of visitors after 2 days of the winter storm.

Note that hyperparameters in the clustering method may influence the clustering results even after we used the elbow method. We tested the number of clusters from 2 to 10 and found that there are generally two patterns in the spatial distribution and differences in the visit count. Getting more clusters would let certain POI categories dominate that cluster, though. Analyzing the specific POI category needs business and other domain knowledge. We left this for future discussion.

The role of heterogeneous information

We elaborate on the heterogenous information captured by our model during the analysis of place resilience beyond visit counts. Unlike previous studies that solely focused on studying differences in visit counts, our model incorporated heterogeneous information, including the category of the points of interest (POI) and visitor travel behavior, to better understand resilience.

We calculated two types of distance measures: (1) the distance between differences in visit counts and (2) the distance between embeddings pairwise for each POI. This approach allowed us to explore the relationships between POIs and their resilience characteristics from two different perspectives. We plotted each POI pair distance as a point on a graph, with the differences in visit counts on the x-axis and the embedding distance on the y-axis, as depicted in Figure 3.

This graph was divided into four distinct areas. If our model had only captured differences in visit counts, the graph would appear as a straight-line crossing regions i) and iii). However, since our model incorporates additional heterogeneous information, we anticipate observing more complex patterns in the graph. The experiment results are shown in Figure 4.

Figure 4.

Differences in the visit count for each cluster. Four solid lines show the median in each cluster.

In the graph Figure 5, areas (ii) and (iv) revealed the most distinct information that our model was able to capture. In area iv), we observed a pair of points representing two POIs: Captain’s Landing and Chartwells. Captain’s Landing, being a very popular restaurant with hundreds of daily visitors, was paired with another POI with a significantly lower visitor count, approximately 20 or less. These two points were located at the bottom right of the graph. Upon mapping them, we identified that they were situated in close proximity to each other.

Figure 5.

Distance comparison. Every node is a pair of POI. The x-axis is the distance between the differences of visit count for each POI pair. The y-axis is the distance based on our embeddings.

Additionally, we found that many visitors to Captain’s Landing and Chartwells came from the same places. This interconnectedness between the two POIs highlights the importance of considering visitor origins and their impact on resilience.

In area ii), the pairs of points had similar visitor counts and resilience levels. However, upon tracing back to the graph, we discovered that the travel distances for these visitors were different. Some visitors were locals, while others came from far away locations. This distinction in travel distances was verified through the original safe graph data. The results align with the findings of (Qiang and Xu, 2020), which indicated that travel distance significantly influences resilience.

Furthermore, the graph exhibited a dense distribution close to the y-axis. This density signifies that even minor differences in visit counts were effectively captured by our model based on other attributes of the POIs.

Discussion

Assessing resilience through heterogeneous graph analysis: A frontier approach

The integration of heterogeneous graph analysis into resilience studies marks a significant advancement in our understanding of complex systems. This approach, rooted in the capabilities of Graph Neural Networks, offers a nuanced perspective into the multifaceted relationships that exist within a disaster, particularly between various places.

Our method excels in detailing the intricate relationships between different places. For instance, consider a community affected by a natural disaster. Traditional models might look at the community as a singular entity, whereas our method can delve deeper, examining relationships between the flow of people and local institutions. This granularity allows us to understand how different components of a community interact and influence one another, especially in the face of adversity. Our approach offers a fresh perspective on resilience studies. Instead of viewing systems as monolithic entities, we advocate for a more granular, relationship-focused approach. This shift in perspective can lead to novel insights and more effective interventions.

These intricate relationships are paramount in resilience studies. They provide insights into the vulnerabilities and strengths of a system. By understanding these relationships, we can predict how a system might respond to shocks, and where interventions might be most effective. Our method, therefore, not only identifies the existing resilience of a system but also potential avenues for enhancing it.

With a deeper understanding of disaster impact, mitigation efforts can be more targeted and effective. For instance, our method identified the visitors were affected differently for various type of POI, then mitigation efforts can focus on customize solutions for each cluster of places. This targeted approach ensures that resources are used efficiently and have the maximum possible impact.

One of the standout features of our method is its adaptability. While our study focused on communities affected by natural disasters, the principles and techniques can be applied to a range of other scenarios, from human-induced disasters to economic shocks. The versatility of our method means that it can offer valuable insights across a range of contexts, highlighting its strengths and pinpointing areas where interventions can be most effective.

Implications for policy and practice

Our approach not only surpasses traditional linear regression and homogeneous graph models in accuracy but also provides a nuanced understanding of how different types of data can be integrated to improve predictive capabilities. It represents a significant shift towards more dynamic and precise resilience assessments. Urban planning agencies should consider incorporating such models into their analytic toolsets. This shift requires policies that support the training of staff and the allocation of resources towards the adoption of graph neural networks. Enhancing computational capabilities within planning departments can ensure that cities are not only reacting to disasters but proactively planning for them with greater precision.

The distinct patterns of visitation changes observed in different clusters during disasters highlight the necessity for dynamic resource allocation strategies. Emergency response could be tailored based on our method to provide support where it is most needed during the various phases of a disaster, improving overall community resilience. Our model’s ability to cluster POIs based on resilience characteristics allows for tailored policy interventions. For instance, different resilience strategies can be designed for areas with a high density of traveler accommodations versus areas dominated by bars and restaurants. This level of specificity in policy design, facilitated by advanced modelling, can lead to more effective allocations of resources before, during, and after disaster events. Urban planners can utilize these insights to draft zoning laws and resource allocation plans that reflect the unique needs of each urban cluster.

The inclusion of demographic, geographic, and categorical data within POI nodes has shown to significantly improve the accuracy of predictions concerning place resilience. Policies should promote the integration of heterogeneous data sources in urban planning and emergency management systems to enhance predictive capabilities and response strategies. Leveraging data from private entities such as businesses and social platforms can enrich public datasets, enhancing the robustness of resilience models. Encouraging public–private partnerships through incentives and supportive legislation can help build comprehensive datasets that are beneficial for both planning and real-time disaster response. The interdisciplinary nature of our research highlights the benefits of integrating insights from diverse fields such as computer science, geography, and social sciences in urban planning. Policies that foster collaboration across academic disciplines and between the public and private sectors can drive innovation in resilience planning. By supporting partnerships and funding interdisciplinary research projects, policymakers can facilitate the development of new methodologies that are both innovative and applicable in real-world settings.

Limitation and future directions

One of the primary limitations is the data quality that impacted by the mobile phone users, sampling strategy from US census, etc. The quality and completeness of the data directly impact the accuracy of the model. Incomplete or biased data could lead to misinterpretations or inaccuracies in the model’s predictions. The reliance on detailed demographic and behavioral data in our model also raises important considerations regarding privacy and security. It is essential for policymakers to enact robust data governance policies that protect individual privacy while allowing for the comprehensive analysis needed to enhance urban resilience. This includes strict adherence to data protection laws, transparent data usage policies, and the implementation of secure data handling and storage practices.

Additionally, while GNNs are powerful tools for modelling complex systems, they can sometimes be difficult to interpret. The “black box” nature of these models can make it challenging to understand exactly how they are making their predictions. For example, we checked if certain visitor CBG’s age and gender impacted the cluster results, but there are no significant statistical differences for each cluster. This phenomenon may be resulted from the measurement limitation of CBG data as we aggregated both POI and CBG into clusters hence individual information cannot be expressed easily. Though the model captures those visitor’s CBG data, it suffers from explain it, which is a major limitation of this method. This is an area where further research is needed to develop more interpretable models.

Future directions for this research could include incorporating more diverse data sources to improve the model’s accuracy and robustness. Additionally, more research could be done to improve the interpretability of GNNs, making them more accessible and useful for decision-makers. Finally, future work could also explore the application of this approach to other types of complex systems, expanding its potential impact.

Conclusion

In this study, we introduced a novel approach using a heterogeneous graph neural network to analyze place resilience. Our method incorporates diverse datasets into resilience modelling, revealing concealed patterns compared to traditional models. Empirical experiments during the winter storm Uri in Galveston, TX, demonstrated a close link between resilience, visitor information, mobility, and place characteristics. Our model unveiled two resilience patterns: immediate impact for some places and delayed effects for others, correlated with points of interest (POI) locations. A key strength is its applicability beyond the studied region and disaster due to the lack of spatial or temporal restrictions. This approach offers insights into resilience factors across different events, aiding policy-making.

Future research can enhance the model’s generalization and explore measures to improve resilience in various places. Analyzing resilience through human mobility helps uncover strengths and vulnerabilities, informing practical solutions and designs for future challenges. Our study contributes to place resilience knowledge, leveraging diverse datasets and advanced graph neural networks to guide decision-making and foster resilient communities.

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the National Science Foundation under 2401860 and 2430700, as well as NASA under 80NSSC22KM0052. The funders had no role in the study design, data collection, analysis, or preparation of this article.

ORCID iDs

Jiaxin Du

Xinyue Ye

Data availability statement

The datasets that were utilized for analysis in the present study were initially accessible through the Safegraph repository (safegraph.com). However, there has been a transition to a different platform known as Advan. The datasets can now be accessed via subscription at .

Jiaxin Du is an assistant professor in the College of Computing at Grand Valley State University. His research expertise is on urban informatics and GeoAI.

Xinyue Ye is a professor in the Department of Landscape Architecture and Urban Planning and director of Center for Geospatial Science, Applications and Technology at Texas A&M University. His research expertise is on human dynamics and urban informatics.

Xiao Huang is an assistant professor in the Department of Environmental Sciences at Emory University. His research expertise is on human environment interaction.

Yi Qiang is an assistant professor in the School of Geosciences at University of South Florida. His research expertise is on spatial-temporal data modeling.

Chunwu Zhu is a doctoral student in the Department of Landscape Architecture and Urban Planning. His research expertise is on urban planning and urban data science.

References

Adger

(2000) Social and ecological resilience: are they related? Progress in Human Geography 24(3): 347–364, Sage Publications Ltd. DOI: 10.1191/030913200701540465.

Amin

Mazumder

Dumler

, et al. (2022) Virtual testbeds for community resilience analysis: state-of-the-art review, consensus study, and recommendations. Natural Hazards Review 23(4): 03122001, American Society of Civil Engineers. DOI: 10.1061/(ASCE)NH.1527-6996.0000582.

Barbosa

Barthelemy

Ghoshal

, et al. (2018) Human mobility: models and applications. Physics Reports 734: 1–74. DOI: 10.1016/j.physrep.2018.01.001.

Berke

Newman

Lee

, et al. (2015) Evaluation of networks of plans and vulnerability to hazards and climate change: a resilience scorecard. Journal of the American Planning Association 81(4): 287–302, Routledge. DOI: 10.1080/01944363.2015.1093954.

Cai

Lam

NSN

Qiang

, et al. (2018) A synthesis of disaster resilience measurement methods and indices. International Journal of Disaster Risk Reduction 31: 844–855. DOI: 10.1016/j.ijdrr.2018.07.015.

Cai

Newman

Lee

, et al. (2023) Simulating the spatial impacts of a coastal barrier in Galveston Island, Texas: a three-dimensional urban modeling approach. Geomatics, Natural Hazards and Risk 14(1): 2192332.

Cutter

Barnes

Berry

, et al. (2008) A place-based model for understanding community resilience to natural disasters. Global Environmental Change 18(4): 598–606, Local evidence on vulnerabilities and adaptations to global environmental change. DOI: 10.1016/j.gloenvcha.2008.07.013.

Zhang

, et al. (2020) Assessing vulnerability of road networks based on traffic flow betweenness centrality: a case study in Wuxi. Journal of Zhejiang University, Science Edition 47(2): 223. DOI: 10.3785/j.issn.1008-9497.2020.02.013.

Dubey

Bryde

Dwivedi

, et al. (2022) Impact of artificial intelligence-driven big data analytics culture on agility and resilience in humanitarian supply chain: a practice-based view. International Journal of Production Economics 250: 108618, Special Issue celebrating Volume 250 of the International Journal of Production Economics. DOI: 10.1016/j.ijpe.2022.108618.

10.

Falco

(2019) Participatory AI: reducing AI bias and developing socially responsible AI in smart cities. In: 2019 22ND IEEE international conference on computational science and engineering (IEEE CSE 2019) and 17th IEEE international conference on embedded and ubiquitous computing (IEEE EUC 2019), New York, NY, 1-3 August 2019, pp. 160–164. DOI: 10.1109/CSE/EUC.2019.00038.

11.

Fey

Lenssen

(2019) Fast graph representation learning with PyTorch geometric. ICLR Workshop on Representation Learning on Graphs and Manifolds. 2019.

12.

Gao

, et al. (2023) Exploring flood mitigation governance by estimating first-floor elevation via deep learning and google street view in coastal Texas. Environment and Planning B: Urban Analytics and City Science 51(2): 296–313.

13.

Haggag

Siam

El-Dakhakhni

, et al. (2021) A deep learning model for predicting climate-induced disasters. Natural Hazards 107(1): 1009–1034. DOI: 10.1007/s11069-021-04620-0.

14.

Hamilton

Ying

Leskovec

(2017) Inductive representation learning on large graphs. Advances in Neural Information Processing Systems 30.

15.

Hao

Wang

(2023) Modeling dynamics of community resilience to extreme events with explainable deep learning. Natural Hazards Review 24(2): 04023013, American Society of Civil Engineers. DOI: 10.1061/NHREFO.NHENG-1696.

16.

Haraguchi

Nishino

Kodaka

, et al. (2022) Human mobility data and analysis for urban resilience: a systematic review. Environment and Planning B: Urban Analytics and City Science 49(5): 1507–1535. DOI: 10.1177/23998083221075634.

17.

Heydari

Huang

Hiraoka

, et al. (2023) Estimating inter-regional mobility during disruption: comparing and combining different data sources. Travel Behaviour and Society 31: 93–105. DOI: 10.1016/j.tbs.2022.11.005.

18.

Holling

(1973) Resilience and stability of ecological systems. Annual Review of Ecology and Systematics 4(1): 1–23. DOI: 10.1146/annurev.es.04.110173.000245.

19.

Dong

Wang

, et al. (2020) Heterogeneous graph transformer. Proceedings of The Web Conference 2020 2020: 2704–2710.

20.

Wang

She

, et al. (2021) Human mobility data in the COVID-19 pandemic: characteristics, applications, and challenges. International Journal of Digital Earth 14(9): 1126–1147.

21.

Huang

Jiang

, et al. (2021) The characteristics of multi-source mobility datasets and how they reveal the luxury nature of social distancing in the US during the COVID-19 pandemic. International Journal of Digital Earth 14(4): 424–442.

22.

Mai

Janowicz

Yan

, et al. (2020) Multi-scale representation learning for spatial feature distributions using grid cells. In: International conference on learning representations, 2020, Addis Ababa, Ethiopia, 30 Apr 2020. https://openreview.net/forum?id=rJljdh4KDH

23.

Malczewski

Jankowski

(2020) Emerging trends and research frontiers in spatial multicriteria analysis. International Journal of Geographical Information Science 34(7): 1257–1282, Taylor & Francis.

24.

Mehta

Pancholi

Saxena

(2024) Organizational resilience and sustainability: a bibliometric analysis. Cogent Business & Management 11(1): 2294513. DOI: 10.1080/23311975.2023.2294513.

25.

Podesta

Coleman

Esmalian

, et al. (2021) Quantifying community resilience based on fluctuations in visits to points-of-interest derived from digital trace data. Journal of The Royal Society Interface 18(177): 20210158.

26.

Popovich

Wallace

Penney

, et al. (2021) Mapping the winter storm’s impact. The New York Times, https://www.nytimes.com/interactive/2021/02/16/us/winter-storm-texas-power-outage-map.html (accessed 29 July 2023).

27.

Qiang

(2020) Empirical assessment of road network resilience in natural hazards using crowdsourced traffic data. International Journal of Geographical Information Science 34(12): 2434–2450, Taylor & Francis. DOI: 10.1080/13658816.2019.1694681.

28.

Qiang

Huang

(2020) Observing community resilience from space: using nighttime lights to model economic disturbance and recovery pattern in natural disaster. Sustainable Cities and Society 57: 102115. DOI: 10.1016/j.scs.2020.102115.

29.

Rosenheim

Guidotti

Gardoni

, et al. (2021) Integration of detailed household and housing unit characteristic data with critical infrastructure for post-hazard resilience modeling. Sustainable and Resilient Infrastructure 6(6): 385–401, Taylor & Francis. DOI: 10.1080/23789689.2019.1681821.

30.

Saad

Hagelaar

van der Velde

, et al. (2021) Conceptualization of SMEs’ business resilience: a systematic literature review. Cogent Business & Management 8(1). DOI: 10.1080/23311975.2021.1938347.

31.

Sculley

(2010) Web-scale k-means clustering. Proceedings of the 19th international conference on World wide web 2010: 1177–1178.

32.

Shin

Yuan

Siong

, et al. (2017) Towards a deep learning powered query engine for urban planning. In: 2017 international conference on Asian language processing (IALP), Singapore, 5-7 December 2017, 99–102.

33.

Song

Newman

Huang

, et al. (2022) Factors influencing long-term city park visitations for mid-sized US cities: a big data study using smartphone user mobility. Sustainable Cities and Society 80: 103815.

34.

Tan

Guo

Mohanarajah

, et al. (2021) Can we detect trends in natural disaster management with artificial intelligence? A review of modeling practices. Natural Hazards 107(3): 2389–2417. DOI: 10.1007/s11069-020-04429-3.

35.

Zhou

, et al. (2022) Resilience assessment and improvement for cyber-physical power systems under typhoon disasters. IEEE Transactions on Smart Grid 13(1): 783–794. DOI: 10.1109/TSG.2021.3114512.

36.

van de Lindt J Ellingwood

McAllister

, et al. (2018) Modeling community resilience: update on the center for risk-based community resilience planning and the computational environment IN-CORE. In: 17th U.S.-Japan-New Zealand workshop on the improvement of structural engineering and resilience (ATC-15-16), Queenstown, NZ, 15 November 2018.

37.

van der Maaten

Hinton

(2008) Visualizing data using t-SNE. Journal of Machine Learning Research 9(Nov): 2579–2605.

38.

Wang

(2018) Social media analytics for natural disaster management. International Journal of Geographical Information Science 32(1): 49–72, Taylor & Francis. DOI: 10.1080/13658816.2017.1367003.

39.

Wang

Shi

, et al. (2019) Heterogeneous graph attention network. The world wide web conference 2019: 2022–2032.

40.

Wang

Zhang

Huang

, et al. (2022) Urban–rural disparity of social vulnerability to natural hazards in Australia. Scientific Reports 12(1): 13665, Nature Publishing Group. DOI: 10.1038/s41598-022-17878-6.

41.

Wang

Huang

She

, et al. (2023) Diverged landscape of restaurant recovery from the COVID-19 pandemic in the United States. iScience 26(6): 106811.

42.

Yabe

Zhang

Ukkusuri

(2020) Quantifying the economic impact of disasters on businesses using human mobility data: a Bayesian causal inference approach. EPJ Data Science 9(1): 36. 1. Springer Berlin Heidelberg. DOI: 10.1140/epjds/s13688-020-00255-6.

43.

Han

, et al. (2023) Developing human-centered urban digital twins for community infrastructure resilience: a research agenda. Journal of Planning Literature 38(2): 187–199.

44.

Yuan

Esmalian

Oztekin

, et al. (2022) Unveiling spatial patterns of disaster impacts and recovery using credit card transaction fluctuations. In: Environment and Planning B: Urban Analytics and City Science. Thousand Oaks, CA: Sage Publications Ltd STM, 23998083221090250. DOI: 10.1177/23998083221090246.

45.

Zhao

Zeng

, et al. (2023) Revisiting the gravity laws of inter-city mobility in megacity regions. Science China Earth Sciences 66(2): 271–281.

46.

Zhu

Zhang

Wang

, et al. (2020) Understanding place characteristics in geographic contexts through graph convolutional neural networks. Annals of the Association of American Geographers 110(2): 408–420, Taylor and Francis Ltd. DOI: 10.1080/24694452.2019.1694403.

47.

Zou

Lam

NSN

Cai

, et al. (2018) Mining twitter data for improved understanding of disaster resilience. Annals of the Association of American Geographers 108(5): 1422–1441, Taylor & Francis. DOI: 10.1080/24694452.2017.1421897.