Toward Resilience: Assessing Retail Location’s Complex Impact Mechanism Using PLS-SEM Aided by Machine Learning

Abstract

Because urban retail faces challenges in sustaining vitality and viability, risking decay in urban centers, retail space resilience (RSR) has become a pressing concern. Retail location presents an opportunity because it aligns with RSR in maximizing store vitality and adopting a long-term perspective. This study uses PLS-SEM to examine the complex impact mechanism of retail location attributes (LAs) on retail space resilience (RSR), based on 304 retail spaces in central Shanghai. LAs and RSR are assessed based on a metrics system, followed by Random Forest for variable selection. An impact pathway framework grounded in key theoretical models is then constructed. The results from the PLS-SEM analysis show that Amenity exerts the strongest direct influence on RSR (β = 0.383), followed by Agglomeration (β = 0.294) and Accessibility (β = 0.291), while the results of the mediation effect further reveal that RSR is primarily shaped by consumers’ trip-chaining behaviors, with agglomeration effects and the spatial interaction model playing secondary roles. Notably, the scale of the retail space negatively affects RSR (β = −0.016), suggesting that large retail centers may be less resilient due to weaker consumer attachment. Overall, our research suggests that consumers’ perceptions and behaviors play key roles in RSR. Based on this insight, this study proposes placemaking strategies aimed at fostering consumer attachment and developing neighborhood-oriented retail nodes aligned with consumers’ preferences.

1. Introduction

Urban retail spaces play a vital role in enhancing urban vitality, promoting economic diversification [1], and fostering social cohesion and urban livability [2], making them a key factor in the sustainable development of city centers. However, in the face of multiple threats—including e-commerce, suburbanization, shifting consumer preferences, and the COVID-19 pandemic—retail spaces have been increasingly experiencing a decline. For instance, town centers and high streets across European cities have witnessed widespread store closures due to declining foot traffic [3,4]. Similarly, in some Asian cities, shopping malls that once attracted significant investment are now frequently shutting down or being resold, turning from valuable assets into liabilities for investors [5,6]. These trends highlight the vulnerability of physical retail spaces, making resilience in urban retail an urgent and pressing topic.

Considering these circumstances, REPLACIS (Retail Planning for Cities Sustainability), a transnational urban research network project conducted between 2009 and 2011 by researchers from four European universities focused on resilience in urban retail and its contribution to urban sustainability [7]. The final report of REPLACIS defines retail resilience as “the ability of different types of retailing at different scales, to adapt to changes, crises or shock that challenge the system’s equilibrium, without failing to perform its functions in a sustainable way” [8]. Even though this definition conceptualizes retailing as an economic activity, it has elevated the academic awareness of building resilience in physical urban retail spaces. Inspired by this perspective, later research framed resilience in urban retail spaces through the following two key dimensions: vitality and viability [7,9,10]. Vitality refers to the ability of retail spaces to attract and engage customers, bringing together diverse public groups through consumeristic acts and thereby fostering urban life and social interaction [11]. Viability, on the other hand, concerns the commercial space’s ability to sustain operations and maintain its business lifespan over time. Together, vitality and viability form a crucial pair of factors for shaping resilience in urban retail environments.

Building on the concept of retail resilience, scholars have explored its intangible drivers from various angles. Some studies have focused on policy, highlighting how flexible and inclusive retail management can strengthen resilience [1,12,13]. Others, informed by consumer culture theory, stressed the need for retail spaces to adapt to evolving consumer expectations and satisfaction-driven strategies [14]. On the other hand, a group of scholars has examined the impact of geo-spatial attributes of retail spaces on retail resilience. Kärrholm (2014) argues that retail activity is inherently tied to its spatial form, with the physical configuration and urban context playing a critical role in resilience [15]. In his research, he cited the concept of spatial resilience, which refers to the ability of a space to maintain its functions and withstand disturbances, which depends on the interdependence between the space and its surrounding environment [16]. Recent empirical studies have shown that spatial location attributes, including Accessibility [17,18], Amenity [19], Agglomeration [20], and Socio-Demographic features [21], demonstrate a strong correlation to resilience in urban retail spaces.

While the existing literature on retail resilience is extensive, only a limited number of studies have approached the topic from a spatial perspective, contextualizing retail resilience within the framework of urban geography. Among those, most have examined only two or three geo-spatial variables, but retail location is inherently multi-dimensional, and various spatial attributes may influence the resilience of brick-and-mortar retail spaces. This study aims to fill this research gap by investigating the complex impact mechanisms of multiple geo-spatial location attributes on urban retail space resilience.

Retail location theory provides a framework to optimize retail space resilience for two reasons. First, its primary objective is to maximize attraction to customers through strategic selection [22,23], which aligns with the resilience objective. Second, coinciding with resilience strategies, retail locations, which are considered high-risk investments, are assessed with a long-term perspective that takes into account not only the temporal market environment but also future changes [24]. Therefore, this research proposes to establish a theoretical linkage between retail resilience and retail location theory by examining how geo-spatial location attributes influence resilience in urban retail spaces.

This study adopts the term retail space resilience (RSR) to refer to the capacity of urban retail spaces to continuously function as areas for retail activity, encompassing both vitality and viability. RSR reflects not only the ability to attract footfall and generate instantaneous vibrancy but also the capacity to sustain operations and commercial relevance over time. Existing research has primarily focused on confirming the influence of individual location attributes on retail space performance. However, according to complexity theories, when multiple factors are simultaneously acting on a subject, the resulting impact mechanisms are often non-linear, interdependent, and dynamic, making it more complex than one-way correlations [25].

Therefore, this study aims to disclose the complex impact mechanism of location attributes (LAs) on retail space resilience (RSR) based on an empirical study in Shanghai. Our research proposes an integrated methodology that combines Random Forest for preliminary variable selection with partial least squares–structural equation modeling (PLS-SEM) for causal pathway testing. First, we develop an assessment metric system for both RSR (as the explained variable) and LAs (as explanatory variables). Random Forest identifies key LA variables with strong predictive power, which are then embedded into a theoretical framework constructed from a literature review. PLS-SEM is subsequently applied to test the hypothesized pathways. By pairing machine learning with SEM, this approach leverages both statistical rigor and predictive accuracy to uncover robust, interpretable mechanisms, offering deeper insights into how spatial attributes and consumer behavior jointly shape retail resilience.

2. Materials and Methods

2.1. Research Area and Sample Data

As the scope of this empirical study is to capture variations in RSR over a long historical trajectory, our research focuses on the following six central districts in Shanghai: Huangpu, Xuhui, Changning, Jing’an, Putuo, and Hongkou. These areas are known as Shanghai’s Concession, which was established by foreign powers such as Britain and France in the 1880s, during which time the concept of the department store was introduced to Shanghai, making it one of the earliest urban retail hubs in China. These districts have since witnessed multiple historical transformations, from the concession era to the founding of the People’s Republic and the era of economic reform. During these transitions, the retail areas in the city’s core, including Nanjing Road, Huaihai Road, and Chenghuangmiao, continued to operate and still do to this day, demonstrating strong resilience [26]. This long-standing and dynamic retail history makes Shanghai’s central districts a highly valuable context for studying urban retail space resilience.

The data collected on retail spaces are classified as “shopping malls” and “department stores” according to DataShanghai (the official opendata platform operated by Shanghai’s municipal government, https://data.sh.gov.cn/), Amap, and Baidu Map api. Retail outlets embedded into other functional entities without autonomous location choices, including grocery or other value-oriented consumption, are excluded from database sourcing. After accessing the data, approximately 500 retail space POIs were collected, and after a validity check on operation status, scale, and function, 304 out of 500 POIs were selected (Figure 1).

2.2. Retail Space Resilience Assessment

This study proposes the concept of retail space resilience (RSR) based on a comprehensive literature review, defining it as the capacity of urban retail spaces to continuously function as containers for retail activity. Specifically, RSR encompasses the following dual notion: fostering vitality and sustaining viability through urban environment and market changes. To capture the two components of RSR, vitality and viability, this study proposes a measurement index for each.

2.2.1. Customer Footfall—Measurement of Vitality

Customer footfall, or foot traffic, measures the vitality of retail spaces and is critical in building resilience in physical retail spaces [27]. Stable customer footfall fosters adaptive social networks and can guide flexible marketing strategies that help retail spaces to keep up with changing consumer habits and recover quickly from shocks [28].

Therefore, footfall data are widely used as indicators of urban vitality. Navigation platforms with a large user base and location information can be used as sources of real-time footfall data [29,30]. Recent studies based in China utilized Baidu Maps location and navigation data to assess footfall [31,32].

From the real-time footfall data, accessed from Baidu Maps Smart-Eye Urban Population Geography Big-Data Platform, a total of 304 valid POIs from the area of the study were selected as the database of our study.

2.2.2. Retail Lifespan—Measurement of Viability

The lifespan of a retail space serves as a direct indicator of its viability, reflecting how long it can adapt to market shifts while maintaining operations. Pickett et al. (2004) emphasized that the resilience of urban spaces lies in their ability to continuously function amid changing material and immaterial conditions—essentially, to “stay in the game” [33]. Many empirical studies have used the decline or closure of retail spaces as inverse indicators of resilience [18,34,35,36]. Following Holling’s (1973) definition of resilience as the ability to maintain system function [37], this study defines lifespan as the total number of years a location has been used for retail, regardless of changes in management or ownership. Since ownership changes are typically influenced by managerial factors rather than location attributes, they are excluded from this analysis.

2.3. Retail Location Attributes Assessment

2.3.1. Theoretical Background

There are four foundational theories of retail location, as follows: Central Place Theory, Hotelling’s Law, Reilly’s Law of Retail Gravitation, and Bid-Rent Theory [22]. These four theories primarily construct the notion of desirable retail location attributes in the success of retail spaces, from the perspective of economic geography. Later on, in 1980s, as shopping centers and consumer culture had prevailed, consumer behavior theories began to identify desirable location attributes from the demand side, focusing on consumer needs and preferences. Among the most relevant are consumer culture theory (CCT) and the theory of trip-chaining behavior.

Building on insights from both retail location theory and consumer behavior theory, this study develops a measurement framework for retail location attributes aimed at building resilience. The framework includes six location attributes and fifteen measuring variables.

2.3.2. Assessment Metric

• Accessibility

Central Place Theory, proposed by Walter Christaller, in 1933, is a foundational retail location theory. He developed a model suggesting that in a hypothetically uniform and evenly distributed market, the most attractive retail location would be at the center of the area, thereby maximizing accessibility for customers [38]. This theory established that the primary attribute of a successful retail location is being accessible to as many people as possible. However, the theory was developed before the widespread motorization of cities. In today’s urban context, consumer travel patterns are far more diverse, involving both vehicles and public transit [39]. As a result, this study proposes three indices to measure the accessibility of retail locations, as follows:

Global Integration: Bill Hillier 1984 proposed the Space Syntax method, which models human activity based on the configuration of urban street networks [40]. Within this framework, Global Integration is used to indicate a location’s centrality, calculated by measuring the topological depth of each space within the entire network [40,41].

In this study, the DepthMapX platform is used to compute Global Integration values for the study area, and the results are shown in Figure 2a.

Access to Public Transit: Studies have shown that in large Asian cities the accessibility of locations largely depends on the integration of public transit facilities [42,43]. Therefore, in our study, the access to public transit is calculated by the number of access points to public transit within 1 km around each retail location. The public-transit access points were retrieved from the Amap api.

Parking Facility: In high-density Asian cities like Shanghai, the growing number of vehicles often poses serious capacity challenges to retail centers. As a result, both the availability of parking facilities and the walking distance to retail centers are critical considerations for evaluating accessibility.

In this study, the number of parking facilities within a 200 m radius of each retail center is used as an additional measure of accessibility for consumers traveling by private vehicle.

2.

Amenity

According to consumer behavior theory, trip-chaining behavior refers to the chain effect between sequential activities, meaning the likelihood of engaging in the next activity depends on the current one. Shopping, socializing, and entertainment have been identified as the activities most commonly involved in trip chaining, as they often occur as intermediate stops within a sequence of trips [44]. This finding is widely applied in consumer behavior research, suggesting that consumers tend to engage in multipurpose shopping trips [45]. The activity site is assumed to include all opportunities which are accessed without additional travel during a single sojourn, it provides scheduling convenience for consumers and, thus, improved the chance of consumers to visit those retail locations [46].

The 2 measuring metrics of Amenity are calculated as follows:

Diversity: This study uses the Entropy Index to measure the diversity of amenities surrounding each retail location [47]. Amenity categories include healthcare, living services, entertainment, culture, municipal and social institutions, sports facilities, education, and tourism. POI data were retrieved from the Amap api. The formula is provided below:

$$H=-{\sum }_{i=1}^n{p}_i\times {\mathrm{l}\mathrm{o}\mathrm{g}}_2\left(p_i\right)$$

(1)

Open-Space Ratio (OSR): Although public open spaces are often not marked as points of interest (POIs) on maps, those near retail locations provide opportunities for unstructured social interaction and health-related activities, creating positive externalities that attract consumers [35,48,49]. In this study, OSR is calculated as the proportion of open space within a retail site relative to its total area and serves as an open space indicator under the broader Amenity dimension.

The formula is provided below:

$$OSR={\displaystyle \frac{Open space area \left({\mathrm{m}}^2\right)}{Site area \left({\mathrm{m}}^2\right)}}$$

(2)

3.

Agglomeration

In retail location theory, Hotelling’s Law holds equivalent significance to Central Place Theory (CPT). It emphasizes the role of Agglomeration in attracting customers, proposing that retailers selling similar products should be located closer to each other in order to share maximum market area [50]. Also, recent empirical research has shown that retail agglomerations exhibit greater resilience against market shocks, as Agglomeration entails spillover effects, such as shared transportation infrastructure, pooled management resources, and enhanced brand visibility [18,19,51]. The 3 measuring metrics of Agglomeration are calculated as follows:

Kernel Density Estimation (KDE): estimates the spatial distribution of data points based on kernels as weights [52], executed on a GIS platform. The calculation steps are the following:

$$K(u)={\displaystyle \frac{1}{\sqrt{2\pi }}}{e}^{-\frac{1}{2}{u}^2}$$

(3)

Calculate for kernel, using the Gaussian kernel formula, then choose the appropriate bandwidth (h), which is the maximum distance between retail locations to be taken in the KDE analysis. In our study, we use a walking distance of 1 km as the bandwidth, h.

$$\widehat{f}(x)={\displaystyle \frac{1}{nh}}\sum _{i=1}^{n}K\left({\displaystyle \frac{x-x_i}{h}}\right)$$

(4)

Finally, estimate the kernel density, whereas n = the number of total data points; x_i = location of the ith point; K = kernel; and h = bandwidth, which was set to be 500 m in our study.

Nearest Neighbor Index (NNI): Reflects a location’s proximity to other retail locations in terms of the Euclidean distance [53]. In this study, the Distance Matrix tool on a GIS platform was used to calculate the mean Euclidean distance from each retail location to all other retail points within the study area.

$${\mathrm{N}\mathrm{N}\mathrm{I}}_i={\displaystyle \frac{1}{n-1}}\sum _{\begin{array}{c}j=1\\ j\ne i\end{array}}^n{d}_{ij}$$

(5)

where n = the total number of retail locations; d_ij = Euclidean distance between location i and j; the summation excludes i = j to avoid calculating the distance from a point to itself.

Point Density: The number of point features (e.g., retail centers) within a specified search radius around each cell in a raster grid. The density is calculated based on the number of points within a 500 m radius of each cell center. The calculation formula is as follows:

$$\mathrm{Point} \mathrm{Density}={\displaystyle \frac{{\sum }_{i=1}^n{w}_i}{A}}$$

(6)

where w_i = the weight of point i (equal to 1 in our study); n = the number of points within the search radius; and A = the area of the search circle.

4.

Scale

In 1953, Reilly proposed a retail gravitation model based on Newton’s law of gravity, suggesting that the scale of retail spaces acts as an attractive force, while distance to consumers functions as a repelling force. The final decision of consumers to visit a particular retail location results from the interplay of this counteractive pair [54]. Then, David Huff (1964) [55] developed the Huff Model to estimate potential market volume. In this model, the positive variable is the retail center’s floor area, which serves as a direct variable of its gravitational pull: the larger the scale, the stronger its attraction to consumers [55]. The 2 measuring metrics of Scale are calculated as follows:

Site Area: calculates the scale of the site of retail locations, measured from OpenStreetMap and a GIS platform;

Floor Area: calculates the scale of the retail space itself, retrieved from DataShanghai, Local Space Viewer, and direct surveys.

5.

Socio-Demography

Early studies in consumer behavior theory recognized that Socio-Demographic attributes of consumers influence their purchasing behavior [56,57]. However, the primary objective of these early theories was to guide pricing strategies and product assortments based on customer profiles [58]. More recent empirical research has shifted focus, emphasizing that residential and employment presence around retail locations plays a crucial role in sustaining those spaces by ensuring consistent foot traffic and fostering retail resilience [21,59]. Therefore, this study proposes two variables to measure the Socio-Demography attribute of retail locations:

Employment: number of companies within 1 km around retail locations, retrieved from the SPDO open-data platform and supplemented with the Amap api;

Residential Settlement: total site area of residential settlement within 1 km around retail locations, retrieved from land-use data on DataShanghai.

6.

Publicity

Consumer Culture Theory (CCT) posits a dynamic relationship between consumer action and the cultural meanings embedded in the marketplace. CCT argues that consumers are motivated by culture and symbolic value, engaging in consumption as a means to construct personal and collective identities [60]. This symbolic value also extends into the physical environment of cities [61]. Consumers tend to associate the physical environment attributes of retail spaces with their taste, status and identity [62]. Therefore, this study defines Publicity as the ability of a retail location to convey its symbolic value to consumers, and incorporates it as a key dimension within the retail location attribute framework.

Visibility: Victor Gruen, the pioneer of the modern shopping center, highlighted the importance of visibility in shaping a retail location’s symbolic identity and consumer perception [63]. This study measures street-level visibility using the isovist property analysis on Depthmap X [64], with a 2 km radius to reflect average visual range (Figure 2b).

Consumer Review: Review scores from digital platforms reflect public perception and the symbolic value of retail locations through collective consumer evaluations.

This study uses the scores of each retail locations on Dianping api, a Chinese independent third-party consumer review platform with the largest database.

Social Media: In the digital era, social media plays a key role in placemaking [65], often shaping consumer perceptions before physical visits [66].

This study measures the Social Media variable by counting Weibo posts pinned to each retail location, using data retrieved from the Sina Weibo API, China’s most widely used platform.

2.4. Research Framework

The structural equation model (SEM) method integrates factor analysis and regression to model multiple impact pathways [67] and is widely used in consumer behavior and retail management studies due to its ability to estimate both direct and mediated effects, especially when factors span across spatial and socio-demographic dimensions. Unlike traditional regression, SEM accommodates interdependent relationships and simultaneous causal chains, providing a more comprehensive understanding of multifactor systems like retail space resilience (RSR).

Random Forest (RF), a decision-tree based machine learning algorithm, effectively identifies the distribution of importance across explanatory variables, represented as feature importance value [68]. It builds multiple classifiers with randomly selected features, evaluating each variable’s importance by permuting values and measuring their effect on prediction error [69].

Therefore, RF can be incorporated with the PLS-SEM method as a preliminary variable selection tool, because it can screen a set of hypothesized influencing variables and retain the ones with higher importance. Prior studies (e.g., Zhu et al., 2024; He et al., 2024) have successfully combined RF and PLS-SEM to select features and analyze structural relationships related to carbon emissions [70,71].

The empirical analysis mainly consists of 3 sections: Data Collection, Theoretical Framework Construction, and the SEM Analysis of the Impact Mechanism, as shown in Figure 3. In the Data Collection section, a database of 304 retail entities in Shanghai is constructed using 15 location attribute (LA) indices and 2 RSR indices (Appendix A,

Table A1. Summary table of the database of Shanghai’s 304 retail space samples, including 2 RSR and 12 LA metrics (Z-score standardization).

Variable	Min	25%	Median	75%	Max
Footfall	−2.23	−0.70	−0.12	0.55	5.48
Lifespan	−1.09	−0.60	−0.19	0.26	5.23
Floor Area	−1.56	−0.86	−0.16	0.89	1.59
Site Area	−1.56	−0.87	0.00	0.87	1.57
KDE	−0.91	−0.91	−0.5	0.32	2.37
NNI	−3.84	−0.60	0.22	0.80	1.55
Point Density	−1.21	−0.77	−0.26	0.37	3.09
Global Integration	−2.4	−0.9	−0.15	0.60	2.09
Public Transit	−1.64	−0.70	−0.19	0.51	7.12
Parking Facility	−2.31	−0.78	−0.01	0.66	3.57
Diversity	−4.15	−0.62	0.11	0.66	3.19
Open-Space Ratio	−2.57	−0.60	−0.03	0.60	3.19
Visibility	−1.74	−0.74	−0.18	0.61	3.75
Consumer Review	−1.56	−0.87	0.00	0.87	1.56
Social Media	−0.57	−0.56	−0.47	0.14	4.86
Residential Area	−4.00	−0.57	0.17	0.76	2.05
Employment	−2.62	−0.69	0.04	0.59	2.76

), based on the assessment metric system detailed in Section 2.3.2.

The theoretical framework’s construction is divided into two part. Firstly, Random Forest is employed for the preliminary variable selection. The model predicts the combined influence of LA variables on RSR, and SHAP (SHapley Additive exPlanations) analysis is applied to interpret each variable’s contribution [72]. Location attribute variables with significant and robust feature importance and SHAP values will be retained while the ones with low or unstable contribution will be screened out. Based on this, a theoretical framework of impact pathways between location attributes and RSR is proposed through a review of the relevant literature. The selected variables from the Random Forest are subsequently integrated into the framework as measuring variables.

Finally, partial least squares SEM (PLS-SEM) is used to test the hypothesized pathways. PLS is particularly suitable for complex models with formative indicators, smaller samples, and non-normal distributions [73]. The path coefficients (β) generated by the model provide empirical evidence of both the magnitude and direction of impact each LA variable has on RSR. In addition to direct pathways, the model also reveals mediating effects, allowing the identification of both primary and secondary influence chains. This integrated approach provides a robust, interpretable structure for understanding how location attributes jointly shape resilience in urban retail spaces.

3. Results

3.1. Database Collection

3.1.1. Explained Variable: Retail Space Resilience

Footfall data, sourced from Baidu Map’s Smart-Eye platform at a 170 × 170 m grid scale, reflect retail space vitality and were retrieved as real-time data on 25 January 2025. Retail lifespan was obtained from online sources or direct inquiries as an indicator of viability. For illustrative purposes, a footfall value was then weighted with the normalized value of the lifespan, as shown in Figure 4b. The lifespan weighted result exhibits a stronger geographical correlation, as shown in Figure 5. Retail spaces with both high lifespan and footfall values are concentrated in central locations like Nanjing Road, Huaihai Road, and Changshou Road, confirming a preliminary link between Retail Space Resilience (RSR) and their locations.

3.1.2. Explanatory Variable: Location Attributes

The explanatory variables consist of six location attributes (LAs), each measured with two indices, totaling 15 variables. Raw geo-spatial data were processed based on the corresponding assessment metric and then visualized using IDW interpolation at a 2 m distance (Figure 5).

3.2. Theoretical Framework Construction

3.2.1. Preliminary Variable Selection Using Random Forest

The Random Forest (RF) model is applied to predict the RSR based on a combined impact of 15 LA variables. Then, the contribution of each LA variable into the whole prediction model are identified based on the feature importance analysis aided with the visualization using SHAP.

Before training, data pre-processing and feature engineering were performed. In our model, the footfall metric for the data samples was weighted by their corresponding lifespan values in order to reflect the RSR, where samples with a longer lifespan weigh more in the whole calculation matrix. The lifespan value was normalized using Z-score standardization.

The dataset was split into training (80%) and testing (20%) sets. Hyperparameter tuning was performed using GridSearchedCV. RF underwent hyperparameter tuning using five-fold cross-validation to evaluate 81 hyperparameter combinations, resulting in 405 total evaluations to determine the best settings, and the optimal parameters are presented in

Table 1. The hyperparameter used in the Random Forest model for predicting the 15 LA attributes’ combined impact on the RSR.

Hyperparameter	Value	Description
n_estimators	100	The model uses 100 trees, ensuring robust learning.
max_depth	10	Maximum depth of each tree that limits how deep the tree can grow
min_samples_split	1	Minimum samples required to split an internal node.
min_samples_leaf	5	Minimum samples required at a leaf node to contain enough data.

.

Then, the RF model was trained and predictions were made on the test set, and the feature importance output is illustrated in Figure 6. This prediction result achieved reasonable reliability with a Train R²: 0.897, Test R²: 0.572, RMSE: 21.584, and MAE: 16.726. As shown in the feature importance map, both diversity and open space ratio within the Amenity attribute of retail locations show substantial feature importance. Kernel density estimation and global integration shows low feature value, meaning that these two variables does not have much influence in the overall combined impact toward RSR. Then the rest of the LA variables shows moderate impact in the test set.

Based on an acceptable R², SHAP plots were generated to compute feature importance. For each individual prediction, SHAP calculates the exact contribution of every LA variable compared to the average prediction, by testing different combinations of all 12 variables. Then, the model computes SHAP values for each LA variable across all data points, and the results are summarized in the beeswarm plot in Figure 7a and encoded by color scale in the heatmap in Figure 7b. The SHAP values can clearly identify variables that strongly influence predictions, help detect unexpected relationships, and make the RF prediction result more transparent.

As shown in Figure 7, amenity diversity and open-space ratio have the widest spreads over the SHAP value and the darkest pink shade in the heatmap, indicating that locational Amenity has the highest positive impact on predicting RSR. On the other hand, as illustrated in both the feature importance value and SHAP value in Figure 6 and Figure 7, the three LA variables—“global integration”, “KDE”, and “consumerReview”—show a low contribution to the combined impact on RSR.

Therefore, based on the two outputs of the Random Forest model—the feature importance and SHAP values—this study screened out three variables from the original 15 location attributes (LAs) due to their feature importance scores falling below 0.02, indicating minimal contribution to the combined impact on RSR. These excluded variables are KDE, consumerReview, and global integration. A subsequent impact mechanism analysis was conducted using the remaining 12 LA variables.

3.2.2. Theoretical Hypothesis

Based on the Random Forest model’s prediction of the combined impact of 15 location attributes (LAs) on retail space resilience (RSR), this study retains 12 key variables to construct a theoretical framework. To guide hypothesis development, three interdisciplinary theories are adopted, as follows: spatial interaction models (SIMs), agglomeration effects theory, and trip-chaining behavior. These frameworks align with the dual nature of RSR, which includes both momentary vitality and long-term viability. Building on these three theoretical models, this section proposes three corresponding modules of impact pathways, which, together, form a comprehensive framework capturing the complex influence of retail location attributes on RSR.

• Spatial Interaction Model

The spatial interaction model rationalizes the flow of people, goods, and economic activities within spaces as a result of decision-making processes. Rooted in assumptions from the social sciences, these models construct mathematical formulations to explain spatial flows and have been widely applied across disciplines. Notably, they are frequently used in consumer behavior research to analyze the likelihood of individuals choosing to visit specific retail locations [74].

One of the key principles within spatial interaction models is distance decay, which posits that as the travel distance between entities increases, the intensity of spatial interaction tends to decrease [75,76]. When applied to the relationship between retail spaces and consumers, this suggests that a consumer’s likelihood of engaging in interaction activities at a retail space is constrained by the physical distance between them.

Accordingly, in SIMs, a place’s accessibility is a positive predictor of spatial interaction [77]. SIMs imply that retail spaces can strengthen their attractiveness to consumers by enhancing physical proximity and improving connectivity to transportation infrastructure. Therefore, our study proposes the following hypothesis:

H1. 

Accessibility positively influences resilience in retail spaces.

In the spatial interaction model developed by Wilson (1971), the scale of a retail facility is also taken into account when estimating its attractiveness to consumers [77]. According to SIMs, the scale of a retail facility delineates its drawing power, indicating that larger facilities have a stronger capacity to attract consumer flows [77], which is in line with Reilly’s Gravitational Law and the Huff Model [55].

Subsequently, Andris, C., increased the understanding of urban interactions by proposing that interaction is essentially a social network embedded in urban locations. Individuals with established social ties tend to patronize the same geographic locations at approximately the same time [78]. In this view, retail spaces function as nodes within these social networks, and the interactions occurring at these nodes can be captured and analyzed using location-based social media data [79,80]. Building on the SIM framework, which identifies Scale and Accessibility as the two main drivers of a retail facility’s drawing power, and incorporating recent advancements that define interaction as social interaction measurable through digital traces, this study further integrates social-media-derived variables into its modeling of retail vitality. Therefore, this study proposes the following hypotheses:

H2. 

The Accessibility of retail locations positively influences Publicity.

H3. 

The Scale of retail locations positively influences Publicity.

• Agglomeration Effects Theory

Agglomeration effects theory (or economies of agglomeration) explains the formation of economic clusters in urban spaces based on economic rationales. The primary driver behind urban agglomeration is the advantage of cost savings [81]. Specifically, the clustering of retail entities can significantly reduce transportation costs by enabling retailers to share the flow of urban consumer traffic—particularly that which is concentrated around public transit stations [82].

Beyond cost efficiency, social networks and public recognition also function as forms of capital, offering additional economic advantages. Agglomerating near highly visible locations reduces transaction and marketing costs for retailers [83]. Saving on both transportation costs and marketing costs helps retail centers to be viable [84].

Such a form of agglomeration strengthens the branding of the retail mix and, thus, increases the attractiveness to consumers [85], as it is easier to create an image of a remarkable shopping destination for the retail clusters. Also, being located near each other creates convenience for consumers to compare different brands, saving time and cost when browsing around different retailers and, therefore, also increases the attractiveness of retail agglomeration for customers [86,87].

Therefore, based on the agglomeration effects theory that is applied in the research field of retail spaces, three hypotheses are proposed:

H4. 

Agglomeration positively influences RSR.

H5. 

Accessibility positively influences Agglomeration.

H6. 

Publicity positively influences Agglomeration.

• Trip-Chaining Behavior

Trip-chaining behavior is a strand of consumer behavior theory that focuses on the rationale behind why consumers commit to multi-purpose shopping trips [45]. As mentioned in the assessment metric system in Section 2.3.2, consumption and socializing tend to happen between a chain of trips. Therefore, consumers tend to shop at locations that provide opportunities for a diverse range of activities, which in turn contributes to the resilience of retail spaces.

Furthermore, as mentioned in the trip-chaining behavior theories, whether the initiative is a household or work-related trip is a decisive factor in whether it will trigger a chain of other events, including consumption. Residential households tend to trigger a chain of interdependent trips [46,88]; however, work-related trips have relatively low chances of inducing a chain of trips because work is mostly the primary and sole purpose of these trips [89]. This suggests that the socio-demographic features surrounding a retail location, classified as residents or employers, determine the chances of customers committing to a chain of various trips, thus influencing the demand for diversified amenities. Residential settlements are hypothesized to be a positive index, while employment is hypothesized to be a negative index.

Therefore, our study proposes the following hypotheses:

H7. 

Amenity positively influences retail space resilience.

H8. 

Socio-demographic features surrounding a retail location influence urban amenity.

3.2.3. Theoretical Framework of Impact Pathways

Combining the pathways based on the theoretical analysis, this article proposed a theoretical framework of retail location attributes’ complex impact mechanism for retail space resilience, as shown in Figure 8. The explanatory variables are the six proposed location attributes, which were measured with the variables screened from the preliminary variable selection; the explanatory variable, RSR, is measured with customer footfall and retail lifespan, representing the vitality and viability of urban retail spaces. This theoretical framework will then be exemplified using the PLS-SEM method in Section 3.3.

3.3. Structural Equation Model Analysis of Complex Impact Mechanisms

The final part of the empirical analysis is to employ the PLS-SEM method to determine the complex impact mechanism of LAs on RSR, based on the data-driven framework proposed in the last section. The PLS-SEM analysis was performed using the SmartPLS 4 analysis program.

3.3.1. Data Validity Testing

This study assesses internal consistency using Cronbach’s α and composite reliability (CR), with all α values > 0.7 and CR values well above 0.7, indicating strong reliability and stability (

Table 2. Validity tests of the location attributes and the measure indices.

Items	Factor Load	Cronbach’s Alpha	CR	AVE
Publicity	-	0.701	0.862	0.759
Social Media	0.974	-	-	-
Visibility	0.724	-	-	-
Scale	-	0.742	0.886	0.795
Floor Area	0.886	-	-	-
Site Area	0.898	-	-	-
Accessibility	-	0.822	0.918	0.849
Parking Facility	0.919	-	-	-
Public Transit	0.923	-	-	-
Socio-Demography	-	0.843	0.927	0.864
Employment	0.923	-	-	-
Residential Settlement	0.936	-	-	-
Agglomeration	-	0.784	0.902	0.821
NNI	0.920	-	-	-
Point Density	0.893	-	-	-
Amenity	-	0.798	0.904	0.825
Diversity	0.949	-	-	-
Open-Space Ratio	0.865	-	-	-

). For convergent validity, all measurement variables are significant and have standardized factor loadings > 0.70 and AVE values > 0.5, confirming good convergent validity. Discriminant validity was tested using the Fornell–Larcker criterion and HTMT, showing lower internal than external correlations and HTMT values < 0.90, verifying discriminant validity. Harman’s single-factor test reveals four factors with eigenvalues > 1, with the first factor explaining 29.31% of the variance, which is below the 40% threshold. The construct correlations are all < 0.9, indicating no severe common method bias. Multicollinearity was checked using VIF, with values between 1 and 3, which is well below 5, confirming no multicollinearity.

3.3.2. Verification of the Theoretical Framework of Impact: PLS-SEM Test Results

This study employed the Bootstrap method with 5000 resamples to analyze the structural model and identify relationships between constructs. The final results and the model are shown in

Table 3. Path coefficients and hypothesis testing.

Impact Path		Path Coefficient (β)	t-Value	p-Value	Effect Size
H1	Accessibility → RSR	0.291	4.995	***	Large
H2	Accessibility → Publicity	0.345	6.922	***	Large
H3	Scale → Publicity	−0.150	3.371	**	Medium
H4	Agglomeration → RSR	0.294	4.511	***	Large
H5	Accessibility → Agglomeration	0.421	5.919	***	Large
H6	Publicity → Agglomeration	0.371	8.328	***	Large
H7	Amenity → RSR	0.383	8.677	***	Large
H8	Socio-demography → Amenity	0.473	9.811	***	Large

** p < 0.01, *** p < 0.001.

and Figure 9. The SEM result’s explanatory power of the explained variable RSR is measured with an R² value of 0.485, which is considered moderate explanatory power in the scenario of PLS-SEM tests [90,91,92], meaning that the test result is sufficiently robust to explain nearly half of the variance in RSR based on the included predictors.

The results indicate that the hypotheses H1 through H8 are supported and meet the acceptance criteria. All pathways tested were found to have highly significant values, as shown in Table 3, and, therefore, indicate a reliable test result. The path coefficients shown in Figure 9, which indicates the strength of the pathways, are distributed evenly in the model, mostly ranging from 0.29 to 0.47, indicating a balanced and systematic impact mechanism. Among the pathways, the mediating impact of Accessibility–Agglomeration and Socio-Demography–Amenity score had the highest pathway coefficients, which were both above 0.4. This indicates that within a well-balanced network of impact, the location attributes demonstrate a relatively high internal level of connection.

The complexity of the model necessitates an analysis of certain intermediate relationships to observe their mediating effects on other constructs. The significance values of the mediation effects between constructs are listed in

Table 4. Mediation effect test.

Impact Path		β	LLCI	ULCI	t-Value	p-Value
H9	Socio-Demography → Amenity → Footfall	0.181	0.127	0.237	6.529	***
H10	Scale → Publicity →Agglomeration → Footfall	−0.016	−0.032	−0.006	2.423	*
H11	Scale → Publicity → Agglomeration	−0.056	−0.094	−0.025	3.116	**
H12	Publicity → Agglomeration → RSR	0.109	0.052	0.168	3.562	***
H13	Accessibility → Agglomeration → RSR	0.124	0.018	0.059	4.489	***

* p < 0.05, ** p < 0.01, *** p < 0.001.

. The percentile confidence interval method was used to test the indirect effects [93]. The results show that the [LLCI, ULCI] values for all paths do not include 0, indicating that the mediation effects of all paths are significant. Hypotheses H9 through H13 are supported. The calculation of the VAF (variance accounted for) further confirms that H9 through H13 exhibit partial mediation effects.

In the tested framework, Scale, Publicity, and Socio-Demographic attributes of retail locations are mediating variables that impact indirectly upon RSR. And Accessibility, Amenity, and Agglomerations are three direct influencing location attributes. Among the five tested mediating impact pathways, H9, H12, and H13 demonstrated stable β coefficients, ranging from 0.109 to 0.181, meaning equivalent significance among these three mediating pathways, while H10 and H11 suggest weaker path strength, with β coefficients of −0.016 and −0.056, indicating that the scale of the retail locations facilities a weak and negative mediating effect on RSR.

What is noteworthy is that the attribute Publicity is involved in both positive and negative impact pathways H10-H11. Based on the positive β coefficients of H2, β = 0.345, and H6, β = 0.371, Publicity can be identified as a positive mediating attribute. However, given that H3 had β = −0.150, Publicity is negatively influenced by Scale and, thus, contributes to a negative impact pathway originating from Scale and, ultimately, affecting RSR.

4. Discussion

4.1. Empirical Validation of Classical Location Attributes and the Emerging Primacy of Amenity in Retail Space Resilience

The SEM results confirm that Amenity, Accessibility, and Agglomeration are three retail location attributes with direct impacts on the resilience of physical retail spaces. Among them, Amenity, measured by the diversity and open-space ratio, exerts the strongest influence (β = 0.383), while Accessibility (β = 0.291) and Agglomeration (β = 0.294) play secondary but still significant roles. These findings empirically validate the relevance of Central Place Theory and Hotelling’s Law in contemporary retail contexts. However, classical retail location theories do not fully recognize the importance of amenity diversity and integration with open spaces in enhancing retail space resilience. This theoretical gap stems from the historical objective of retail location theory, which focused on maximizing immediate foot-traffic volume. In contrast, the objective of this study is resilience, which incorporates not only present vitality but also long-term operational viability. The results demonstrate that a diverse surrounding mix of amenities and the integration of open spaces significantly enhance a retail space’s ability to retain both vitality and viability over time.

These results can be interpreted through the lens of fundamental resilience theory, which was originally rooted in the ecological discipline and introduced by Holling in 1973 [94]. This theory emphasizes the capacity of ecological systems to absorb shocks and continue functioning without collapsing. Later, sociologists extended the concept to socio-ecological systems, highlighting that the ability of social systems to regenerate and reorganize fundamentally relies on an environment with more diversified functions [95]. Recent empirical studies have supported this view in the urban retail context. During the lockdown periods of the COVID-19 pandemic, retail locations situated near a diverse mix of public amenities, which helped compensate for disrupted daily functions, demonstrated stronger resilience [96]. Amenity diversity enhances retail resilience by attracting consumers through greater convenience, a wider variety of activities, and a richer environmental experience [97].

Furthermore, this study includes non-programmed open spaces as a distinct dimension of amenities, which not only demonstrated measurement validity in the model but also showed a significant positive impact on retail space resilience. The underlying reason lies in the ability of integration with open spaces within retail centers to create spatial diversity in experience. Pine and Gilmore (1999) introduced the concept of the “experience economy”, emphasizing that the experiential quality of space is a key factor in attracting consumers [98]. Increasingly, retail managers and scholars recognize that enclosed indoor environments alone can no longer meet consumer demand for experience-driven consumption. This has led to the emergence of “demalling” strategies—reintroducing features of outdoor shopping streets into retail centers [99]. By developing open spaces, retail centers can expand their functions beyond traditional commerce to include strolling, relaxing, and socializing, ultimately increasing dwell time and enhancing consumer engagement [100]. Moreover, the integration of retail spaces with urban open spaces enables better morphological and social integration with the city fabric, strengthening their long-term viability and community relevance [101].

Although the results of this study validate classical central place theories and spatial interaction models, indicating that Accessibility positively influences RSR, this study places greater emphasis on the role of Accessibility facilities, including public transit infrastructure for pedestrians and parking facilities for private vehicle users, in enhancing retail resilience. In contrast, via Random Forest, the traditional notion of centrality proposed by classical Central Place Theory was shown to be a relatively weak contributing factor in this context. This suggests that, in contemporary urban contexts, consumer preferences tend to prioritize the convenience and quality of access infrastructure, such as multi-modal transit connections and parking availability, over the traditional emphasis on spatial proximity or reduced travel distance. This shift reflects a broader transformation in mobility patterns and lifestyle expectations within modern cities, where the accessibility of a retail location is no longer defined solely by centrality or Euclidean distances but by the ease, flexibility, and comfort with which consumers can reach a destination.

4.2. Constraints of Agglomeration Effects and the Growing Influence of Consumer Trip Chaining

In the impact mechanism framework constructed in this study, Publicity and Accessibility are identified as the primary drivers of agglomeration effects, both of which have been confirmed to be strong impact pathways. This indicates that the agglomeration patterns of retail spaces in Shanghai are mainly shaped by the presence of accessibility facilities (H5 β = 0.421), followed by the influence of social network accumulation (H6 β = 0.371). In the central districts of Shanghai, there are 11 nodes combining metro stations and retail complexes, each with daily passenger flows ranging from 30,000 to 170,000 people [102]. Since centralized transit nodes generate substantial and stable customer footfall, retailers prefer clustering in surrounding areas [103,104].

While the underlying mechanisms of retail agglomeration were validated, SEM analysis reveals that consumer trip-chaining behavior (H9 β = 0.181) exerts a stronger influence on retail space resilience than agglomeration effects (H12 β = 0.109; H13 β = 0.124). This suggests that in Shanghai, resilience by urban retail spaces is primarily driven by consumers’ multi-purpose trip behaviors, whereas the economic advantages brought by agglomeration are secondary. An underlying reason for this is that, nowadays, retail development has gradually transitioned from producer-oriented to consumer-centered [105]. Early retail location theories were developed during the formative stages of physical retail, when retail activity was largely production-driven. However, in today’s context, consumers have access to multiple consumption channels, including e-commerce, which has fundamentally shifted the power dynamic in favor of consumers [106]. Moreover, the benefits of the agglomeration effect in urban retail are not unlimited. In highly saturated and homogenized retail environments, including those increasingly observed in Shanghai, excessive agglomeration may intensify direct competition among similar retailers and even accelerate the elimination or marginalization of less competitive retail units [107].

Moreover, classical trip-chaining behavior theory primarily associates consumption with households and residential settlements and does not fully recognize the role of individuals as employees engaging in retail consumption. This study expands upon the traditional perspective by demonstrating that not only do residential functions have a positive relationship with resilience in retail spaces, but employment density in surrounding areas is also positively linked to consumer behavior. The underlying reason is that a rich offering of urban amenities attracts high-quality human capital to central business districts, where innovative employment opportunities are concentrated. This dense and affluent workforce contributes to a strong consumer base, thereby enhancing the economic resilience of nearby retail spaces [108,109].

4.3. Limited Influence of Mall Scale: Rethinking Spatial Interaction Models in Modern Retail Environments

Finally, this study reveals a point of tension with the classical spatial interaction model, particularly regarding the relationship between retail space scale and spatial interaction. Traditional theories posit that larger retail facilities exert greater drawing power and, thus, attract stronger spatial interactions. However, the impact mechanism analysis in this study shows that retail scale exhibits a weak and even negative mediating effect on RSR, suggesting that a larger size does not necessarily translate into greater resilience in today’s urban retail contexts.

In fact, criticism of the large-scale retail-complex model emerged over a decade ago, with scholars highlighting issues such as poor accessibility, parking shortages, and an oversupply of internal functions as key drawbacks of oversized retail spaces [110]. More recently, researchers have pointed out that large-scale retail centers typically adopt standardized, top-down management strategies—including uniform advertising, publicity, and tenant mix planning—that marginalize the voice of individual retailers, effectively reducing them to employees of the mall. This lack of flexibility hampers a mall’s ability to adapt to evolving consumer preferences, ultimately undermining its resilience [111]. Moreover, the standardized atmosphere of large malls significantly weakens personalized customer relationships. In a study by Gilboa et al. (2024), most customers described their spatial experience in large malls as a “sensory overload” and reported an absence of place attachment and belonging [112]. Other studies similarly observed a lack of warm interpersonal relationships within large chain malls [113]. This emotional detachment is reflected in this study’s SEM results, which show a negative correlation between Scale and Publicity, indicating that larger malls may be less effective in cultivating symbolic and emotional connections with consumers.

Empirical evidence supports these findings. A 2011 report by the UK-based Local Data Company (LDC) noted that larger retail centers exhibited higher vacancy rates and weaker sales performance compared to small- to medium-sized retail spaces [114]. Similarly, Enoch et al. (2020) found that following the COVID-19 lockdown, footfall in six large-scale UK shopping centers dropped by 57–75%, whereas smaller centers showed much more stable vitality levels [115]. These findings underscore the instability of consumer attachment to large shopping centers and further question their long-term viability in fostering retail resilience.

However, as shown in Table 4, the p-value of the mediating effect of scale on RSR is approximately 0.05, indicating relatively weak statistical significance compared to other mediating effects. This suggests that the impact of scale on retail space resilience is not consistently negative and may vary depending on specific contextual factors. In fact, a study on shopping centers in Madrid found that large-scale expansion, when accompanied by the introduction of richly diverse lifestyle experiences, was viewed as a deliberate strategy to enhance resilience. The advantage of large retail centers lies in their ability to integrate a wide variety of consumer demands within a single venue, including children’s play areas, skating rinks, and even indoor ski slopes. In other words, when a large scale is combined with uniqueness and diversity, it can still embody resilience and generate strong market appeal [116].

5. Conclusions

By combining lifespan with footfall in assessing retail space resilience, this study elaborated on traditional retail location objectives from immediate customer footfall attraction to long-term viability. Then, based on empirical results in Shanghai, we bridged two fields of retail resilience and retail location by delineating the complex impact mechanism between them. PLS-SEM confirmed causal links and identified Amenity (β = 0.383), Agglomeration (β = 0.294), and Accessibility (β = 0.291) as direct influencing locational attributes, with Socio-Demography, Scale, and Publicity playing mediating roles in agglomeration effects and consumers’ trip-chaining behavior. The test results refine the theoretical hypothesis by recognizing the primary contribution of consumers’ perceptions on retail space resilience.

Our study contributes to the application of the PLS-SEM method in academic research by proposing a hypothesis framework based on a preliminary feature selection using machine learning, instead of relying solely on theoretical hypotheses. A Random Forest model integrated with the SHAP method was used to sort out the importance of each assessed location attribute variable and retain the location attributes that are involved within the combined impact toward retail space resilience in the subsequent SEM analysis. This combined approach is validated by the high statistical significance of the PLS-SEM test results.

The findings of this study reveal that consumer behavior and perception of retail locations have become the primary anchors of retail space resilience today. Based on this, the study recommends developing neighborhood-oriented retail nodes aligned with the daily trip chains of residents and workers, which are home–commute–shopping or home–leisure–shopping patterns, to increase the likelihood of spontaneous shopping behavior. For instance, in areas with high employment density or office clusters, mixed-use markets combining convenience, placemaking, and open-office functions can be introduced. In residential areas, participatory neighborhood markets could be established to encourage community engagement. This approach shifts retail spaces from being sole-purpose destinations to structural nodes within consumers’ daily paths.

On the other hand, this study confirms the effectiveness of agglomeration effects within Shanghai. Retail spaces benefit from clustering in locations with high social visibility and strong transportation connectivity, which reduces operational costs and fosters positive spillover effects. To preserve these advantages while avoiding the risks of homogeneity-driven competition, this paper recommends the establishment of a “Retail Cluster Quality Assessment”, which evaluates retail agglomeration entities in terms of diversity, functional complementarity, and market segmentation, aiming to prevent the excessive repetition of similar retail formats and reduce systemic competitive risks. Furthermore, this study proposes a “Differentiated Operation Incentive” to guide retail agglomeration. This would involve tax incentives or rental subsidies to attract retail formats that serve different consumer segments and paces of life, which could refer to mixing casual social-oriented businesses with faster-paced service-oriented ones. Such strategies would enhance the resilience of the retail agglomeration landscape by increasing internal diversity and adaptability.

The limitation of our empirical study mainly lies in the particularities of Shanghai’s urban context and consumer behaviors, making the observed results potentially grounded in this specific context. Therefore, even though the insights are theoretically valuable, their direct empirical applicability to other cities with different socio-economic dynamics may be limited.