2.1. Retail location

The selection of retail store locations has long been recognized as a critical determinant of retail success. Classical spatial theories laid the foundation for understanding retail distribution and consumer attraction. Central Place Theory [21] and Reilly’s Law of Retail Gravitation [22] emphasized the roles of accessibility, store attractiveness, and spatial equilibrium in shaping retail patterns. Building on these ideas, Huff’s probabilistic trade-area model further highlighted consumer catchment and probabilistic store choice [23], while subsequent analog models [24], spatial interaction models [25], and econometric approaches [26] became widely adopted tools for predicting store performance and optimal siting. Together, these approaches established a robust theoretical basis for explaining how distance, market size, and surrounding competition influence retail location outcomes.

Retail location research has since evolved from classical spatial explanations toward increasingly data-driven and behaviorally nuanced approaches. Later studies extended the focus beyond distance and market potential, showing that store performance also depends on format differentiation, competitive structure, and local urban context [10,27]. In recent years, the retail landscape has diversified, marked by the proliferation of e-commerce and the rapid rise of “new retail” or online-to-offline (O2O) formats [11]. These developments have fundamentally influenced the criteria and mechanisms underlying retail location choice. Rather than diminishing the importance of physical location, the integration of digital channels, advanced logistics, and data-driven consumer insights has introduced new layers of complexity to location strategy [28,29]. This shift does not only affect new entrants; it also pressures established, traditional retail to adapt and refine their business models and location strategies. At the same time, it is worth noting that traditional retail itself is not a static category. With socio-economic transformation, demographic shifts, and evolving consumer preferences, the internal composition and operational focus of traditional retail have undergone notable changes [4,30]. Store formats have diversified, operational scales have become more polarized, and location strategies are increasingly differentiated according to capital investment and market positioning [6,16]. Existing studies have begun to reveal substantial heterogeneity across retail formats—for example, between specialty grocers and traditional supermarkets, or between new retail and conventional offline retail—suggesting that location logic varies systematically with business model, capital intensity, and target market [3,6,7]. For instance, research comparing Starbucks and Luckin Coffee has demonstrated that new retail brands adopting O2O models exhibit unique sensitivities to demographic, socioeconomic, and competitive factors compared to more established traditional retail [3,11]. In parallel, there is growing attention to the heterogeneity within traditional retail itself, with studies contrasting the location strategies of specialty grocers and traditional supermarkets in Western contexts [6]. These works reveal that even within brick-and-mortar retail, different subtypes, distinguished by product range, target demographics, and service models, respond differently to their urban and competitive environments.

However, despite these advances, there remains a significant gap in the literature regarding the spatial strategies of traditional retail differentiated by their operational scale and asset intensity. This is a critical omission, as the scale and asset structure of retail formats fundamentally shape their location logic, resilience to market changes, and roles in urban spatial structure. Consequently, it is necessary to systematically compare the location strategies of different traditional retail formats, particularly light -asset and capital‑intensive types.

2.2. Determinants of retail location in the urban context

A comprehensive array of objective factors has been shown to shape retail location choice. Demographic variables such as population density, income, education, household size, and age structure are consistent predictors of site attractiveness and sales performance [3,15–17]. Socioeconomic determinants, including rental costs, land prices, and economic activity, are particularly salient for capital-intensive retail formats [6,31]. Accessibility remains a foundational principle, with successful sites often characterized by proximity to transport nodes, arterial roads, and public transit [11]. The density and number of surrounding facilities, such as other retail outlets, offices, schools, and health amenities, have gained increased attention as markers of the vibrancy and drawing power of a locality [3,11]. Besides, recent research underscores the competitive environment as a critical factor.

However, beyond these objective environmental determinants, human perceptions of the environment, including factors such as perceived safety, beauty, and wealth, are increasingly recognized as influential in shaping retail location success [11]. Early urban design research demonstrated that perceived environmental quality is strongly associated with walkability and place experience [32]. With the development of street-view imagery and computer vision, urban perception can now be quantified at scale, including perceived safety, visual quality, and emotional impressions of streetscapes [33–35]. In the Chinese context, related studies have also begun to use large-scale street-view datasets to measure urban greenery and visual quality, confirming the value of human-scale environmental indicators in high-density cities [36,37]. While such subjective factors have been extensively studied in consumer behavior and environmental psychology [32,38], their explicit integration into traditional retail location models remains limited. Most existing empirical work continues to focus on quantifiable, external variables, overlooking the ways in which individual and collective perceptions mediate the relationship between urban context and retail performance. This omission is notable given recent evidence that environmental cognition and behavioral preferences can significantly modify both consumer flows and the effectiveness of retail siting [11].

Methodologically, the vast majority of classic literature relies on linear models, which may oversimplify the often complex and nonlinear effects of environmental determinants. The recent adoption of machine learning techniques, such as XGBoost, has enabled researchers to capture nonlinear patterns [3]. Recent advances in explainable artificial intelligence for tree-based models have further made it possible to move from predictive performance to interpretable understanding of feature contributions and interactions [39]. Integrated analytical pipelines combining XGBoost, SHAP, PDP and Bayesian Networks have proven effective for evaluating retail suitability while incorporating streetscape perceptual metrics, offering a solid methodological foundation for unpacking retail location mechanisms [40]. Related geospatial analytics focusing on retail density thresholds further highlight the value of bootstrap-based uncertainty testing for reliable identification of tipping points across multi-source spatial datasets [41]. In addition, studies focusing on commercial data analytics have demonstrated the great potential of large language models for mining heterogeneous commercial text information [42], which delivers new inspiration for enriching multi-dimensional input features in grid-scale retail spatial analysis. The integration of multi-source geospatial data, including POIs, mobile signaling, and socioeconomic statistics, further enhances the explanatory power of location choice models, particularly in new retail scenarios characterized by high consumer mobility and behavioral heterogeneity.

Despite these advances, several important gaps remain in the existing literature. First, although prior studies have compared retail subtypes such as new versus traditional retail or specialty grocers versus conventional supermarkets, few have systematically examined heterogeneity within traditional retail through the lens of asset intensity and operational scale. As a result, the distinct locational logics of light-asset and capital-intensive formats remain insufficiently understood. Second, while recent studies have begun to quantify human perception using street-view imagery and computer vision, such perceptual dimensions are still rarely integrated with objective urban-environment indicators in traditional retail location models. Third, current machine-learning applications in retail geography have mainly focused on prediction, variable importance ranking, or the identification of nonlinear effects, but have seldom provided an integrated framework capable of jointly revealing interaction effects, hierarchical dependencies, and probabilistic structural pathways. Fourth, much of the emerging evidence on perception-aware and data-driven retail location analysis is derived from new retail scenarios or non-Chinese contexts, leaving a relative lack of empirical evidence from high-density Chinese cities where intense competition, rapid mobility, and mixed urban functions may reshape retail locational mechanisms. Addressing these gaps, this study develops an integrated framework combining XGBoost, SHAP, PDP, ISM, and BN to compare light-asset and capital-intensive traditional retail in Shenzhen, and to examine how objective urban conditions and human perception jointly shape their spatial density patterns.

3.1. Analytical framework and study area

The analytical framework for this study is illustrated in Fig 1. It combines multi-source urban data collection, systematic variable selection, advanced machine learning modeling, and structural mechanism analysis to reveal the spatial logic underlying the location competition between light-asset and capital-intensive traditional retail formats.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. The analytical framework.

https://doi.org/10.1371/journal.pone.0358597.g001

The empirical analysis is conducted in Shenzhen, a leading metropolis in southern China renowned for its rapid urbanization, diverse retail ecosystem, and pronounced spatial heterogeneity. Shenzhen represents an ideal case for examining retail location strategies due to its high population density, dynamic socioeconomic environment, and the coexistence of both light-asset and capital-intensive traditional retail. For spatial granularity, the study area is partitioned into uniform grid units. To mitigate potential bias caused by the modifiable areal unit problem, we implemented a spatial resolution sensitivity test with three uniform grid specifications (300 m, 500 m, and 800 m). We re-ran the full XGBoost-SHAP-PDP-ISM-BN workflow for each grid size and cross-compared variable importance rankings, critical threshold values and hierarchical association structures. The results show highly consistent spatial patterns, feature importance hierarchies and critical threshold intervals across the three grid resolutions. Quantitatively, pairwise Spearman’s rank correlation coefficients of SHAP variable importance rankings between different grid sizes all exceed 0.99 for both retail formats; detailed comparative variable importance tables, Spearman correlation values and bootstrapped threshold estimates with 95% confidence intervals are provided in S1–S6 Tables of the Supporting Information, verifying the robustness of core findings to spatial unit granularity. The 500-meter grid was ultimately selected as the primary analytical unit because it aligns with the typical walking distance for daily retail consumption in Chinese cities and balances sample volume with fine-grained spatial detail. This resolution captures subtle variations in the urban environment well while maintaining adequate statistical reliability in the retailer distribution data, consistent with widely adopted spatial analysis standards in existing urban studies [43]. In the baseline analysis, grid cells lacking either light-asset or capital-intensive traditional retail POIs were excluded to maintain comparability across the two retail categories. We further address the potential selection bias introduced by this filtering rule through a hurdle-model-based robustness check, as detailed in Section 4.2. In this study, convenience stores are taken to represent light-asset formats, while large-scale supermarkets and shopping malls are chosen as representatives of capital-intensive formats. Retail location data are derived from the 2024 Gaode Map POI dataset, ensuring high accuracy and contemporaneity. After filtering, the final sample comprises 3,326 valid grid cells for light-asset retail and 893 for capital-intensive retail, as shown in Fig 2.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. The study area.

(a) Light-asset traditional retail: 500m grid Cells in Shenzhen. (b) Capital-intensive traditional retail: 500m grid cells in Shenzhen. Base map data from OpenStreetMap, available under the Open Database License (ODbL).

https://doi.org/10.1371/journal.pone.0358597.g002

3.2. Data and variables calculation

Based on relevant literature and the characteristics of traditional retail [6,11], this study specifies two dependent variables corresponding to light-asset and capital-intensive formats, and a set of independent variables capturing six dimensions of location determinants: demographics, urban function, competition environment, traffic location, accessibility, and human perception. Table 1 summarizes the definitions, calculation methods, data sources, and data acquisition time for all variables. All measurements were conducted at the 500 m grid level using multi-source spatial data from Gaode Map, Baidu Map, Baidu Huiyan population heat map, and OpenStreetMap for the year 2024.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Variable definitions and calculation method.

https://doi.org/10.1371/journal.pone.0358597.t001

3.2.7. Human perception.

We incorporated six perceptual attributes, including vegetation, beautiful, depressing, lively, safe, and wealthy, to capture the human perception of street environments influencing retail location choice. These indicators represent residents’ perceptions of streetscapes, reflecting greenery coverage, aesthetic appeal, safety, vibrancy, affluence, and overall emotional tone. Evidence suggests that such perceptions significantly affect customer behavior and retail performance [11,38]. The streetscape feature extraction and analysis process are shown in Fig 3. We obtained 205,000 street-view images from Baidu Maps for Shenzhen. Vegetation was quantified using the Mask2Former segmentation framework trained on the Mapillary Vistas dataset, enabling accurate extraction of physical greenery. Emotional perceptions were derived from a ResNet‑152 model trained on the Place Pulse 2.0 dataset, producing scores for the five attributes.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Analytical framework for human perception evaluation.

Base map data from OpenStreetMap, available under the Open Database License (ODbL).

https://doi.org/10.1371/journal.pone.0358597.g003

To mitigate cross-cultural bias introduced by the Western-dominated Place Pulse 2.0 training dataset, we conducted a localized perception validation survey covering 15 representative street blocks across Shenzhen. These blocks were selected via stratified random sampling covering five major urban functional typologies (residential, central commercial, industrial, mixed-use and old urban village districts) to capture heterogeneous streetscape conditions across the city. A total of 80 local residents aged 18–65 completed anonymous on-site questionnaires. All five perceptual attributes (beautiful, lively, safe, depressing, wealthy) were evaluated using a 0–100 continuous visual analogue scale (VAS), matching the 0–100 output range of the ResNet152 perception model to enable direct numerical comparison. For correlation testing, all individual questionnaire responses were aggregated to generate block-level average human perception scores for each of the 15 sample blocks; we similarly averaged all street-view image model predictions within each corresponding block, and Pearson correlation coefficients were computed on these paired block-level aggregated values. Pearson correlation analysis demonstrates a significant positive correlation (r = 0.76, p < 0.001) between crowd-sourced local survey scores and model-generated perception values. Dimension-specific correlation results are as follows: beautiful (r = 0.81, p < 0.001), lively (r = 0.78, p < 0.001), depressing (r = 0.75, p < 0.001), safe (r = 0.73, p < 0.001), wealthy (r = 0.71, p < 0.001). All dimension-level correlations remain strongly statistically significant. This confirms that the pre-trained model’s outputs can reliably reflect Shenzhen residents’ subjective street evaluations despite its Western training base.

All on-site questionnaire procedures involving human participants followed institutional research governance standards. The participants were informed of the study objectives before the experiment and provided written informed consent. This study was approved by the Harbin Institute of Technology Medical Ethics Committee under ethics approval number HIT-2026049.

3.3. Data processing

3.3.2. Shapley Additive exPlanations (SHAP).

SHAP is a unified framework for interpreting the output of machine learning models by quantifying the contribution of each feature to individual predictions [48]. Rooted in cooperative game theory, SHAP values are derived from the concept of Shapley values, which assign a fair distribution of a total payout to players based on their marginal contributions across all possible coalitions.

In the context of machine learning, each feature is treated as a “player” in a coalition, and the model prediction is considered the “payout.” The SHAP method computes the average marginal contribution of a feature by considering all possible subsets of features, thereby capturing the interactions and dependencies among features. This approach ensures that the feature attributions are consistent, locally accurate, and additive.

Formally, the explanatory model is articulated as:

(1)

Here, refers to the individual features under investigation in the next section. The vector describes all possible combinations (coalitions) of the features, where each element indicates whether a feature is included. The feature importance, , denotes the Shapley value attributed to feature .

SHAP facilitates comprehensive interpretability by enabling global insights through feature importance rankings aggregated across the dataset, and local interpretability through visualization tools such as summary plots, dependence plots, and force plots. These visualizations help to reveal complex, nonlinear relationships between features and model outputs, as well as threshold or interaction effects that traditional linear methods may overlook.

3.3.3. Partial Dependence Plots (PDP).

PDP serve as a visualization tool to assess the marginal impact of individual features on the predictions of machine learning models [49]. This method clarifies the associations between variables by means of a partial dependence function, formulated as follows:

(2)

Each variable is associated with a unique partial dependence function , which computes the average value of the model output . Here, is held constant, while varies according to its marginal distribution (). It is important to note that both the model output and the marginal distribution () involve inherent uncertainties.

In this context, represents the partial dependence function for variable , which estimates the expected value of . when is fixed and fluctuates across its marginal distribution (). Given the uncertainties in both and (), we use Monte Carlo integration to approximate Eq. (2) (Eq. 3):

(3)

This approximation utilizes feature values {,  ..., } extracted from the training dataset.

For multiclass classification tasks, PDP employs a one-vs-rest strategy, generating distinct plots for each class. While one-way PDP effectively reveals average marginal effects, it may mask localized interaction effects between variables. To capture these hidden complexities and explore multivariate relationships, we extend PDP to two-way analysis. This approach involves selecting pairs of variables related to the dependent variable and applying computational procedures consistent with Eq. (3) to quantify their joint effects.

Threshold positions along partial dependence curves were objectively identified via change-point detection for automatic extraction of slope inflection points. Bootstrap resampling was conducted at the grid-cell level with replacement over 1,000 replicates. The pre-trained XGBoost model was fixed for all bootstrap iterations; only partial dependence functions were recalculated on each resampled dataset to construct 95% confidence intervals for all estimated threshold breakpoints.

3.3.5. Bayesian Network (BN).

A Bayesian Network (BN) is a probabilistic graphical model that represents a set of variables and their conditional dependencies through a Directed Acyclic Graph (DAG) [50]. In the graph, nodes correspond to random variables and directed edges denote dependency relationships. The statistical influence between connected variables is encoded in conditional probability tables (Fig 4).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Bayesian Network structure and conditional probability table.

https://doi.org/10.1371/journal.pone.0358597.g004

Given variables , ,  ..., , the joint probability distribution factorizes according to the BN’s DAG structure:

(4)

where represents the set of parent variables of .

Bayesian inference updates beliefs in light of new evidence via Bayes’ theorem:

(5)

Here, denotes a hypothesis (e.g., a specific state of a target variable), is observed evidence, is the prior probability, is the likelihood, and is the posterior.

In practice, continuous variables are discretized into non-overlapping intervals before network construction to meet the discrete CPT requirements. Once the structure and parameters are defined, BN enables forward and backward reasoning, sensitivity analysis, and probabilistic prediction of complex systems.

4.1. Comparative spatial patterns of light-asset vs. capital-intensive traditional retail in Shenzhen

Fig 5 and 6 demonstrate distinct spatial distribution patterns between light-asset traditional retail (convenience stores) and capital-intensive traditional retail (large-scale supermarkets and shopping malls) across Shenzhen. Light-asset traditional retail is widely dispersed throughout the city, with the highest average density observed in Bao’an (10.22 stores per grid) and Luohu (9.65 per grid), while peripheral districts such as Longhua and Longgang also maintain relatively high averages (9.63 and 8.61 per grid, respectively). Their notable presence in both central and peripheral areas reflects strong adaptability and low entry barriers. In contrast, capital-intensive traditional retail shows a markedly clustered distribution, primarily located in economically developed districts like Luohu, Nanshan and Futian, where average densities exceed 1.9 per grid. As illustrated in Fig 5(b), over 24.31% of the 500-meter grids host upwards of 13 convenience stores, underscoring the prolific presence of these compact, light-asset traditional retail that thrives through sheer numbers and dense clustering. In stark contrast, Fig 6(b) reveals that a vast majority, 91.16% of such grids accommodate no more than three large-scale supermarkets or shopping malls, entities that, while substantially larger in scale and capital investment, maintain a far more restrained spatial footprint.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Spatial distribution and density of light-asset traditional retail: (a) Light-asset traditional retail in different administrative districts; (b) Density of light-asset traditional retail in Shenzhen.

Base map data from OpenStreetMap, available under the Open Database License (ODbL).

https://doi.org/10.1371/journal.pone.0358597.g005

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Spatial distribution and density of capital-intensive traditional retail: (a) Capital-intensive traditional retail in different administrative districts; (b) Density of capital-intensive traditional retail in Shenzhen.

Base map data from OpenStreetMap, available under the Open Database License (ODbL).

https://doi.org/10.1371/journal.pone.0358597.g006

The observed patterns can be attributed to the differing operational models: light-asset traditional retail requires smaller spaces and lower investments, allowing it to penetrate a wide range of urban contexts and meet daily consumer needs. Capital-intensive traditional retail, on the other hand, depends on larger customer catchments and substantial capital input, which limits its spatial presence to commercial hubs and high-demand areas. Urban planning policies and land use regulations further reinforce these distributional differences.

4.2. Visualization and comparison of variable contributions

To evaluate the determinants influencing the spatial density of light-asset and capital-intensive traditional retail, we tested multiple models including OLS and XGBoost. As shown in Table 2, XGBoost yields an R² of 0.68 for light-asset retail and 0.42 for capital-intensive retail, versus OLS R² values of 0.64 and 0.36. Meanwhile, the MAPE metrics reach 74.51% for light-asset retail and 45.49% for capital-intensive retail under XGBoost. It should be noted that MAPE tends to generate inflated values for count-type outcome variables. The grid-level retail density data contains numerous near-zero observations, which create tiny denominators and artificially magnify percentage error; accordingly, these raw MAPE figures should not be interpreted as direct evidence of weak model fit. Judging purely from absolute numerical standards, the capital-intensive R² is modest and both models carry elevated MAPE levels; nevertheless, XGBoost still surpasses OLS across all R², RMSE, MAE and MAPE benchmarks for both retail formats. Within the context of urban retail geospatial analysis, this level of model performance can still provide useful analytical insight. Retail density is jointly driven by quantifiable built-environment indicators and unmeasured micro-commercial attributes that cannot be fully extracted from open geospatial datasets. Although the model does not fully explain all variation in retail counts, the subsequent SHAP, PDP, ISM, and BN analyses focus primarily on relative variable importance, nonlinear response patterns, and hierarchical associations among predictors. These relational patterns may still offer meaningful and reasonably robust insights even under moderate overall model fit, thereby supporting the credibility of our comparative findings across the two retail categories. To address the potential sample selection bias caused by removing grid cells with zero retail POIs in the baseline model, we further constructed a two-stage XGBoost hurdle model covering the full set of 500 m grid samples (including all zero-density units). This framework separates two independent modeling tasks: a binary XGBoost classifier to predict the probability that a grid hosts retail outlets, and a truncated XGBoost regressor to estimate retail density only for grids with positive retail counts. We replicated the full SHAP variable importance ranking and extracted core nonlinear threshold effects based on hurdle model outputs. The results show highly consistent factor hierarchies, identical dominant influencing variables, and nearly overlapping critical threshold intervals compared with our baseline XGBoost model. This cross-verification confirms that excluding zero-retail grids does not distort the core comparative patterns between light-asset and capital-intensive retail derived in this study. Full performance metrics for the classification and truncated regression stages of the hurdle model, alongside side-by-side SHAP variable importance rankings against the baseline XGBoost, are documented in S7–S10 Tables of the Supporting Information, which numerically demonstrate high consistency in factor hierarchies and dominant influencing variables. Building on the calibrated XGBoost outputs, we applied SHAP to the trained XGBoost models to quantify and visualize the contribution of each variable to the prediction outcomes, enabling a fine-grained understanding of variable importance and directional impacts.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Comparative analysis of model effectiveness.

https://doi.org/10.1371/journal.pone.0358597.t002

SHAP analysis revealed distinct influential factors for the two traditional retail types (Fig 7). For light-asset traditional retail, the three most positively impactful variables were lodging facilities’ density (SHAP = 1.88), non-large supermarkets’ density (1.82), and population density (0.78). These factors likely reflect the preference of light-asset traditional retail for areas with strong service infrastructure and moderate consumer traffic facilitating frequent but low-capital transactions. Conversely, the main negative influencers included general markets’ density (1.47), vegetation (0.82), and medical facilities’ density (0.72), possibly indicating that highly competitive or areas with dense greenery are less attractive for such retail. For capital-intensive traditional retail, medical facilities’ density (0.14), lodging facilities’ density (0.13), and beautiful (0.1) emerged as the top positive contributors. Negative influences for this retail included population mobility (0.22), safe (0.1), and local choice (0.07). Capital-intensive traditional retail tends to choose locations that are well-equipped and aesthetically pleasing but not overly “corridor-like,” as these areas can bring stable customer flow while enhancing the consumer experience and investment returns.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. SHAP-derived feature contributions to retail location density: (a1) Global importance for light-asset traditional retail, (a2) Local importance for light-asset traditional retail, (b1) Global importance for capital-intensive traditional retail, and (b2) Local importance for capital-intensive traditional retail.

Urban functional and competitive variables are more prominent for light-asset retail, whereas perceptual variables are more prominent for capital-intensive retail.

https://doi.org/10.1371/journal.pone.0358597.g007

Examining variable categories further elucidates these patterns (Table 3). For light-asset traditional retail, the highest cumulative relative importance lies in urban function (31.18%), competition environment (27.64%), and human perception (16.42%), with competition environment (13.82%) leading average relative importance. In contrast, capital-intensive traditional retail prioritizes human perception (28.95%), urban function (27.56%), and accessibility (15.79%) cumulatively, while demographic factors (7.79%) top average importance. Notably, human perception exerts a strong impact of on both retailer types, highlighting the critical role of environmental perceptions in shaping traditional retail spatial density. Urban function also significantly influences both types, as mixed-use urban spaces support diverse customer flows and sustainable demand. However, competition environment greatly affects light-asset traditional retail due to their need for agility in competitive markets, whereas demographic stability and accessibility are vital for capital-intensive traditional retail to ensure long-term investment returns.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Variable relative importance for density of light-asset vs. capital-intensive traditional retail.

https://doi.org/10.1371/journal.pone.0358597.t003

Moderate to strong linear correlations are observed between conceptually related paired predictors, including accessibility indicators, transit metrics and emotional dimensions. The full Pearson correlation matrix is provided in S11 Table. Specifically, absolute correlation magnitudes peak at 0.42 for transit variables, 0.72 for the space-syntax Global choice–Local choice pair, and 0.71 for the perceptual Beautiful–Wealthy indicators. While XGBoost maintains stable predictive accuracy under multicollinearity, interrelated predictors will split total SHAP attribution values between one another. For this reason, SHAP importance metrics of variables belonging to the same correlated cluster should be interpreted as collective group contributions instead of discrete independent effects.

4.3. PDP analysis based on XGBoost model

To examine nonlinear associations and joint response patterns between key urban-environment variables and the predicted spatial density of the two retail formats, we conducted one-way and two-way partial dependence analyses for the most influential variables and selected theoretically relevant variable pairs identified from the SHAP results. In Fig 8, cooler colors indicate combinations associated with lower predicted retail density, whereas warmer colors indicate combinations associated with higher predicted retail density. The stability of the identified transition points was evaluated using 1,000 bootstrap resamples. All reported threshold cutoffs can be contextualized against the full empirical distribution of variables via S12 Table in Supporting Information. To improve readability, the main text reports only the principal nonlinear patterns and representative thresholds; the complete set of threshold estimates, interaction intervals, directional changes, and 95% bootstrap confidence intervals is provided in S13 Table of Supporting Information.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 8. Interaction effects of urban environmental variables on traditional retail density derived from PDP analysis.

(a) Light-asset traditional retail. (b)capital-intensive traditional retail. The results reveal moderate-density facility and competition synergies for light-asset traditional retail, but stronger threshold-dependent effects of population mobility and perceived environmental beauty for capital-intensive traditional retail.

https://doi.org/10.1371/journal.pone.0358597.g008

For light-asset traditional retail, lodging-facility density and non-large-supermarket density show generally positive marginal associations with predicted retail density, whereas vegetation shows an overall negative association. General-market density displays a clear nonlinear pattern: its marginal association is positive at low to moderate levels but becomes negative after reaching approximately 31 facilities per grid. This transition suggests that moderate concentrations of general markets may be associated with agglomeration and complementary-demand benefits, whereas higher concentrations correspond to market saturation and intensified local competition.

The two-way PDP results further show that the association between lodging-facility density and light-asset retail density becomes stronger when non-large-supermarket density or population density is also relatively high. General-market density and population density similarly reinforce the favorable association of non-large supermarkets with predicted retail density. By contrast, vegetation exhibits context-dependent joint patterns: although it has an overall negative marginal association, its relationship with retail density varies according to accessibility, population concentration, and the surrounding facility mix. Additional interval-specific patterns involving medical facilities and other service variables are reported in S13 Table of Supporting Information.

For capital-intensive traditional retail, lodging-facility density and medical-facility density show generally positive marginal associations with predicted retail density. Population mobility, however, exhibits an inverted-U pattern. Its association with retail density is favorable up to an intermediate range of approximately 900–1000, after which the marginal effect becomes negative. This finding indicates that moderate mobility is associated with a broader potential customer base, whereas highly transient population patterns may be less compatible with the stable catchment requirements of capital-intensive retail.

Perceived beauty also displays a threshold-dependent relationship. Its marginal association with capital-intensive retail density becomes positive only after the index reaches approximately 35.5. Two-way PDP results further indicate that higher levels of perceived beauty are jointly associated with higher predicted retail density when combined with population mobility, medical facilities, lodging facilities, or transit provision. Service-facility and transit variables likewise exhibit reinforcing joint patterns, suggesting that the concentration of capital-intensive retail is associated with the coexistence of accessibility, facility support, and environmental quality rather than with any single condition in isolation.

Other variables, including perceived liveliness, local street-network choice, lodging facilities, and perceived safety, show additional context-dependent interaction patterns. In particular, the negative conditional association involving perceived safety should not be interpreted as evidence that unsafe environments favor capital-intensive retail. It may instead reflect the broader land-use and morphological characteristics of locations that receive high safety scores but have relatively limited commercial intensity. The detailed interaction intervals and uncertainty estimates for these variables are presented in S13 Table of Supporting Information.

Overall, light-asset retail is characterized primarily by nonlinear competition and service-facility interactions, whereas capital-intensive retail displays stronger threshold-dependent associations with population stability and perceived environmental quality. These thresholds represent model-based empirical transition ranges identified within the Shenzhen sample. They should not be interpreted as causal effects, deterministic site-selection rules, or universally applicable planning cutoffs.

4.4. Model configuration and structure based on ISM and BN

Integrating ISM and BN enables a deeper understanding of the complex, hierarchical relationships among urban environmental variables, offering actionable insights into how these factors collectively drive the spatial dynamics of two traditional retail formats.

Based on SHAP analysis, fourteen key factors influencing the density of both light-asset and capital-intensive traditional retail were identified and subjected to ISM analysis. For density of light-asset traditional retail, the selected factors S₁ to S₁₄ represent lodging, non-large supermarkets, general markets, vegetation, population density, medical facilities, workplace, distance to subway, transit stops, global integration, beautiful, safe, global choice, and population mobility. For density of capital-intensive traditional retail, the variables X₁ to X₁₄ represent population mobility, medical facilities, lodging, beautiful, transit stops, safe, lively, local choice, wealthy, distance to subway, depressing, global integration, non-large supermarkets, and work place.

The adjacency matrices (Tables 4 and 5), which were constructed based on expert validation, encode the directly confirmed influences between factors (1 for presence, 0 for absence). These matrices serve as the basis for constructing the reachability matrices (Tables 6 and 7), which indicate whether an element can reach another, directly or indirectly, within the system (1 for reachable, 0 otherwise). Hierarchical partitioning (Tables 8 and 9) was performed by iteratively comparing the reachability and antecedent sets of each factor, extracting hierarchical levels until the structure for all elements was established.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Adjacency matrix of variables influencing density of light-asset traditional retail.

https://doi.org/10.1371/journal.pone.0358597.t004

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. Adjacency matrix of variables influencing density of capital-intensive traditional retail.

https://doi.org/10.1371/journal.pone.0358597.t005

Download:

• PNG
  larger image
• TIFF
  original image

Table 6. Reachable matrix of variables influencing density of light-asset traditional retail.

https://doi.org/10.1371/journal.pone.0358597.t006

Download:

• PNG
  larger image
• TIFF
  original image

Table 7. Reachable matrix of variables influencing density of capital-intensive traditional retail.

https://doi.org/10.1371/journal.pone.0358597.t007

Download:

• PNG
  larger image
• TIFF
  original image

Table 8. Intersection analysis of reachable and antecedent sets for density of light-asset traditional retail.

https://doi.org/10.1371/journal.pone.0358597.t008

Download:

• PNG
  larger image
• TIFF
  original image

Table 9. Intersection analysis of reachable and antecedent sets for density of capital-intensive traditional retail.

https://doi.org/10.1371/journal.pone.0358597.t009

For density of light-asset traditional retail, the hierarchy consists of six levels: (1) density of light-asset traditional retail; (2) distance to subway, global integration, beautiful, safe, global choice, and population mobility; (3) vegetation, non-large supermarkets, general markets; (4) lodging, transit stops; (5) medical facilities, work place; (6) population density. For density of capital-intensive traditional retail, the structure includes five levels: (1) density of capital-intensive traditional retail; (2) beautiful, safe, lively, local choice, population mobility, wealthy, distance to subway, depressing, global integration; (3) non-large supermarkets; (4) lodging; (5) workplace, transit stops, medical facilities. The ISM hierarchy illustrates the pathways through which urban environment variables affect retail density.

To ensure computational efficiency and scale invariance, variables were discretized using the Jenks natural breaks method. All continuous predictors were split into three quantile-based non-overlapping intervals via Jenks natural breaks binning. Each BN directed acyclic graph was fully constrained by the hierarchical digraph generated from ISM, without any data-driven structure learning. Conditional probability tables for all nodes were estimated using empirical frequency counts from the discretized grid sample dataset. BN inference results are acknowledged to be sensitive to discretization binning schemes, and all pathway interpretations below account for this limitation. Spatial datasets for both retail types were processed in GeNIe for model training. Each Bayesian network model comprises 15 nodes, each discretized into three states, with state probabilities derived from historical data, including prior and conditional probabilities. As shown in Fig 9, integrating the results from GeNIe-based diagnostic analysis, influence path analysis, and sensitivity analysis, two primary pathways were identified as having the strongest impact on density of light‑asset traditional retail. The first pathway operates predominantly through perceptual factors: increased vegetation influences perceived environmental qualities such as safe and beautiful, which subsequently stimulate the clustering of light‑asset traditional retail. This effect can be attributed to the fact that aesthetically pleasing and perceived-safe environments not only attract pedestrian flows but also encourage longer dwell times, thereby increasing the potential customer base and improving the commercial viability of such retailers. The second pathway functions through objective urban environment factors: higher population density increases the density of transit stops, reducing the average distance to metro stations and improving accessibility. Enhanced accessibility facilitates greater consumer mobility and convenience, which are critical for the frequent and spontaneous visits typically associated with light‑asset traditional retail formats.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 9. BN-based evaluation model for density of light-asset traditional retail.

(a) Diagnostic analysis. (b) Influence path analysis. (c) Sensitivity analysis. The results reveal multiple entry points and interconnected pathways underlying light-asset retail clustering.

https://doi.org/10.1371/journal.pone.0358597.g009

As depicted in Fig 10, for density of capital‑intensive traditional retail, in addition to the direct influence of subjective factors such as beauty, safety, liveliness, wealthiness, and the absence of depressing features, a prominent objective factor pathway was identified. Higher medical facility density promotes the growth of lodging facilities, which in turn increases the density of non‑large supermarkets, subsequently influencing population mobility. This chain of effects ultimately fosters an environment conducive to capital‑intensive traditional retail. The underlying mechanism lies in the complementary relationship between these facilities: medical centers and lodging attract steady flows of visitors and citizens, supermarkets provide immediate consumption options, and together they generate sustained foot traffic and diversified demand, both essential for the economic feasibility of high‑investment retail operations.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 10. BN-based evaluation model for density of capital-intensive traditional retail.

(a) Diagnostic analysis. (b) Influence path analysis. (c) Sensitivity analysis. The results reveal a relatively sequential and hierarchical pathway underlying capital-intensive retail clustering.

https://doi.org/10.1371/journal.pone.0358597.g010

The hierarchical stratification derived from ISM clearly distinguishes exogenous core associational factors and intermediate mediating variables for two retail formats. For light-asset retail, population density sits at the bottom foundational layer, with multiple parallel intermediate mediating paths (service facilities, street perception) leading to retail density, which matches its scattered, multi-source agglomeration characteristics. By contrast, capital-intensive retail presents a single sequential upstream chain rooted in medical and lodging facilities, reflecting its reliance on complete supporting commercial clusters before large-scale investment entry. The Bayesian network further quantifies the probabilistic strength of each ISM hierarchical pathway through conditional probability tables. The integrated ISM-BN framework does not merely output static correlation hierarchies; it realizes dynamic scenario reasoning by simulating changes of single or multiple urban environment variables, which provides quantifiable planning reference rather than qualitative description alone. Notably, BN pathway inference is sensitive to the discretization binning scheme adopted here, and all reported association pathways are interpreted with this limitation in mind.

5.1. Nonlinear, interactive, and hierarchical mechanisms shaping spatial density of traditional retail

This study advances the understanding of retail location dynamics by demonstrating that the spatial patterns of light‑asset and capital‑intensive formats are shaped not by linear correlations, but by nonlinear responses, synergistic interactions, and hierarchical dependencies among urban environment variables. The superior performance of XGBoost over OLS confirms that conventional linear models used in previous research overlook critical threshold effects and context‑contingent shifts, potentially leading to misinterpretation of locational logic [11,51].

A key insight is that the same environmental factor can act as either an enabler or a constraint depending on retailer format and local saturation level. This format-based divergence in threshold and interaction mechanisms can be systematically interpreted through classical retail location theories including Central Place Theory and retail agglomeration differentiation theory. The core internal driver lies in fundamental gaps in capital input, operation cycle, consumer positioning and catchment scale between the two retail types. Light-asset convenience stores are short-cycle, low-cost daily retail targeting scattered resident spontaneous consumption, thus forming low-competition tolerance thresholds and weak dependence on streetscape aesthetics. By contrast, capital-intensive supermarkets and malls feature long depreciation cycles, huge land and decoration expenditures, and target comprehensive, planned consumption; they therefore impose strict perceptual quality cutoffs and demand stable resident groups instead of highly mobile transient populations. Competitive density produces positive spillovers for light‑asset retailers up to a moderate threshold, after which congestion effects erode viability, in line with prior studies [52]. Similarly, population mobility also has a format-specific effect: capital-intensive retail benefits from moderate mobility but becomes less strongly associated with locations characterized by excessive population turnover; excessive mobility undermines the operational stability prized by capital‑intensive ventures, thus extending and deepening the conclusions of relevant study [3]. These nonlinear tipping points highlight the need to frame planning interventions around “functional ranges” rather than monotonic assumptions.

Interaction effects detected through two‑way PDP analysis reveal that retailer clustering is seldom the additive result of single variables. Light‑asset formats profit from combinations of functional service proximity and perceptually pleasant surroundings, which jointly lengthen dwell time and stimulate impulse purchases. Capital‑intensive formats, by contrast, require convergence between stable service anchors such as medical, lodging and high‑quality aesthetic environments before density gains emerge, suggesting a dependency on both structural stability and brand‑aligned imageability. Such findings refine classical retail location theory by showing that operational models condition how built and perceived attributes are filtered and prioritised.

The ISM‑BN integration further reframes retailer density as the outcome of structured, multi‑layered correlative chains. For light‑asset formats, clustering can be initiated via multiple “entry points”, either high accessibility through transit and population cores or perceptually enhanced environments that stimulate footfall. This adaptability contrasts with the sequential, dependency‑heavy pathways of capital‑intensive formats, which often hinge on the prior accumulation of supportive anchors before complementary facilities and high‑quality consumer flows can materialise. In both cases, deep‑layer spatial enablers such as accessibility and functional diversity operate indirectly, channelling influence through mid‑tier perceptual and competitive mediators.

By articulating these systemic architectures, the present study moves the debate beyond identifying “which factors matter” to explaining “how and under what conditions they matter.” For urban planners, this implies that optimising retail vitality is less about maximising any single metric and more about orchestrating the timing, balance, and co‑occurrence of supporting conditions. Policies that respect nonlinear thresholds, foster productive facility synergies, and sequence infrastructure, perception, and competition in format‑appropriate ways are better positioned to sustain diverse and resilient retail ecosystems.

5.2. Human perception as a core driver of traditional retail location choice

While the urban environment’s influence on retailer location choice has been widely acknowledged, prior research has predominantly emphasized objective urban features [3,6]. Far less systematic attention has been devoted to understanding how micro-scale street environments, and the combined effects of objective and subjective human perceptions, jointly shape the spatial density of different traditional retail. This study addresses this gap by integrating both dimensions, drawing on multi-sourced urban datasets and advanced computational modeling.

Our findings underscore that human perception metrics are not peripheral, but central determinants of spatial patterns in both light‑asset and capital‑intensive traditional retail, albeit with differing modes of influence. SHAP results reveal that perception‑based indicators account for 16.42% of cumulative importance in light‑asset formats and nearly 29% in capital‑intensive formats, ranking as the most influential category for the latter. This magnitude exceeds or rivals that of purely objective factors, indicating that how an area is experienced can be as decisive for retailer agglomeration as its functional accessibility or demographic base.

The XGBoost‑PDP analysis further confirms that perceptual attributes exert clear threshold effects. For capital‑intensive retail, perceived beauty becomes positively associated with retail density only after a relatively high level of environmental quality is reached. Below this threshold, perceptions may deter investment, even if objective infrastructure is favourable. Light‑asset retail shows a different pattern: perceptual factors such as beautiful or safe do contribute, but their greatest value lies in synergistic interactions with functional variables.

The ISM–BN modelling further clarifies the probabilistic association pathways linking perception and retail density. For light‑asset formats, perception acts as an intermediate amplifier: creating an environment where pedestrians feel safe and visually stimulated not only attracts more flows but also prolongs dwell time, which is particularly valuable for frequent‑purchase, convenience‑oriented businesses. For capital‑intensive formats, perceptions operate more directly as top‑tier nodes in the urban environment hierarchy, shaping market attractiveness even before accessibility or facility mix come into play. This reflects their strategic emphasis on long‑term brand equity, aesthetic alignment, and stability of return on investment.

Domestic urban studies utilizing Chinese street view imagery have advanced quantitative measurement of street greenery and multi-dimensional perceptual aesthetics, and preliminarily linked perceived street quality to urban vitality [36,37,53]. Nevertheless, existing research mostly concentrates on general spatial patterns of perceptual indicators across cities, without unpacking heterogeneous threshold effects of human-centered street perceptions for distinct retail formats. While prior scholarship uniformly aggregates all offline retail outlets for analysis, this study further differentiates light-asset and capital-intensive traditional retail, revealing divergent sensitivity thresholds to street beauty, liveliness and green coverage. It therefore supplements targeted comparative empirical evidence missing from domestic retail location research, and extends street-view-based perceptual analysis to format-specific commercial spatial research.

For urban planners and policy-makers, these results suggest that strategies aimed at enhancing retail vitality cannot rely solely on improving physical infrastructure or demographic targeting. Instead, they must also engineer and sustain favourable perception environments, by investing in streetscape aesthetics, public safety measures, and programming that fosters urban vibrancy. For light‑asset retail, such interventions can magnify the returns of functional clustering; for capital‑intensive retail, they may be a prerequisite for location viability. Integrating perception metrics into planning models thus provides a more complete and predictive framework for sustaining diverse retail ecosystems. From the perspective of retail format segmentation theory, this study supplements a neglected dimension of asset intensity division to existing perceptual location research. Prior retail geography studies mostly treat physical retail as a unified category and ignore that streetscape perception’s threshold constraints are format-specific. Our comparative results complement Central Place Theory by revealing that low-order retail only takes perceived environment as auxiliary attraction, while high-order large-scale commercial facilities regard high aesthetic scores as a necessary precondition for site selection, further enriching the micro environmental interpretation system of classical retail spatial theories.

5.3. Policy implications

The results support a format-specific and threshold-sensitive approach to retail planning. Rather than maximizing commercial density, accessibility, population flow, or environmental quality independently, planning interventions should consider whether these conditions fall within locally appropriate ranges and whether complementary urban functions are jointly present. The empirical transition ranges identified in this study can support local diagnosis and scenario evaluation, but they should not be treated as universal or deterministic regulatory cutoffs.

For light-asset retail, neighborhood planning should support moderate commercial agglomeration while avoiding excessive concentrations of homogeneous outlets. Convenience stores benefit from proximity to residential populations, lodging functions, workplaces, transit services, and complementary small-scale businesses, but excessive local competition may reduce the viability of additional stores. Planning should therefore emphasize functional complementarity and tenant diversity rather than simply increasing outlet numbers. Preserving fine-grained street-front premises and affordable small commercial spaces can also help maintain neighborhood accessibility, grassroots entrepreneurship, and the adaptive capacity of light-asset retail.

Streetscape interventions for light-asset retail should balance environmental quality with commercial visibility and pedestrian access. Greening can improve thermal comfort and walking experience, but its design should avoid obscuring storefronts, interrupting pedestrian connections, or reducing usable commercial frontage. The relevant implication is not to restrict urban vegetation, but to coordinate greenery, frontage design, signage visibility, and pedestrian circulation in neighborhood commercial streets.

For capital-intensive retail, site evaluation should consider the joint presence of stable demand, supportive service facilities, multimodal accessibility, and high-quality public environments. High pedestrian or population mobility alone does not necessarily provide the stable catchment required by large supermarkets and shopping malls. Similarly, streetscape improvements may have limited commercial effects when basic transport connections, facility networks, and market demand are insufficient. Large commercial land releases should therefore be preceded by integrated assessment of catchment stability, surrounding functions, accessibility, and experiential environmental quality.

The hierarchical associations identified by the ISM–BN analysis also suggest that retail-supporting interventions should be coordinated and appropriately sequenced. In prospective commercial hubs, public authorities should align service facilities, transport provision, public-space improvements, and commercial land supply rather than relying on isolated infrastructure projects. However, the model does not imply that a single fixed development sequence will be optimal in every context; the appropriate order should be calibrated to local land-use conditions, existing facility networks, and market demand.

At the citywide level, a dual-track policy framework can help balance inclusiveness and economic competitiveness. Fine-grained commercial spaces should be retained in residential and mixed-use neighborhoods to support light-asset retail and everyday access to essential goods. At the same time, suitable sites with stable catchments, sufficient infrastructure, and high-quality public environments should be reserved or regenerated for capital-intensive retail. Because the numerical thresholds identified in Shenzhen may vary with urban morphology, data resolution, consumer behavior, and local market conditions, their application in other cities requires local recalibration and additional validation.

5.4. Research limitations and future research directions

Four structural limitations merit acknowledgment. First, the framework models location patterns primarily as a function of urban environmental factors, without explicitly incorporating potential strategic interactions among retailers such as first‑mover deterrence and coalition siting. Such game‑theoretical positioning data remain commercially sensitive and are inaccessible to public research, necessitating confidential, long‑term collaboration with industry actors to reconstruct authentic decision chains. Secondly, the ResNet-152 model for street perception relies on the Place Pulse 2.0 dataset, which primarily captures Western urban streetscapes. Although we performed localized questionnaire calibration to align model outputs with Shenzhen residents’ subjective experience, regionally optimized visual perception models trained on Chinese street-view imagery would further eliminate cross-cultural deviations in follow-up research. Thirdly, this study’s variable pool does not fully incorporate several institution- and market-based core determinants that exert decisive impacts on retail layout, including block-level land prices, urban zoning regulations, commercial lease terms, private developer investment strategies, and historical land-use trajectories. Such institutional and commercial data are not publicly available at a citywide scale in Shenzhen, and unified open geospatial databases lack standardized records of these indicators. Future work could cooperate with local urban planning bureaus and real estate developers to integrate these omitted predictors and further enhance model explanatory power. Fourth, although the integrated XGBoost–SHAP–PDP–ISM–BN framework improves the interpretability of complex retail-location mechanisms, some uncertainty in the ISM–BN reasoning process remains unavoidable when modelling urban systems. In this study, such uncertainty was reduced through a two-round Delphi-based expert elicitation process, inter-rater consistency testing, and internal consistency checks; however, the ISM structure still represents a simplified abstraction of inter-variable relationships, while BN inference remains probabilistic and may be affected by data quality, discretization, and finite-sample estimation. Future research could further improve the robustness of the ISM–BN framework by incorporating larger longitudinal datasets, data-driven structure learning, and dynamic Bayesian networks to better capture temporal variation and reduce reliance on expert-defined relationships.