LightGBM–SHAP-Based Study of the Threshold and Synergistic Effects of Physical and Perceptual Scene Elements on Spatial Vitality in Historic Cultural Districts

Abstract

The revitalization of vitality in historic cultural districts can enhance a city’s cultural attractiveness and promote the upgrading of the urban cultural industry and sustainable development. Revealing the threshold and synergistic effects of different districts’ scene elements on district vitality helps to identify the distribution patterns of district vitality and provides a basis for managerial decision-making. This study first uses a geographic information system (ArcGIS) to overlay Baidu heatmaps with the street-network distribution in order to depict the spatiotemporal heterogeneity of district vitality and to compute vitality values by partitions at the district scale. Subsequently, based on an explanatory framework that integrates the physical space and subjective cognition, multi-source data such as street-view panoramas and points of interest (POIs) are quantified to obtain scene-element values for each unit area. Then, the scene-element values and vitality values are integrated into a consolidated database. Additionally, the LightGBM model and the SHAP method are employed to evaluate each element’s marginal contribution and relative importance to district vitality, thereby screening out the key scene elements. Finally, by means of SHAP dependence plots and interaction-effect analysis, the threshold intervals of the key elements and their synergistic relationships are identified, revealing the nonlinear threshold effects and synergies by which scene elements influence spatial vitality. The results show that during rest days, district vitality exhibits stronger diffusion, and the synergistic effect between Leisure-Facility Attractiveness and Street-Network Accessibility is the most prominent in enhancing vitality. High Exhibition-Facility Attractiveness is difficult to sustain crowds on its own; only when Leisure-Facility Attractiveness is likewise high does its effectiveness increase significantly. When Transport Accessibility is within the 0.20–0.40 interval, the positive effect of Leisure-Facility Attractiveness is significantly amplified. An excessive Traditional–Modern Facility Mix readily leads to homogenization of districts; therefore, when introducing modern business formats, local cultural characteristics must be retained. Overall, the generation of district vitality relies more on the synergy between material factors and subjective cognition than on improvements to any single element. The findings of this study provide suggestions for the planning of scene elements and the enhancement of vitality in historic cultural districts.

1. Introduction

Urban development has shifted from incremental expansion to stock optimization [1]. In this context, culture-led renewal of historic cultural districts has become an important driver of urban growth [2]. Studies show that revitalizing the vitality of historic cultural districts not only strengthens residents’ cultural identity and sense of belonging but also enhances a city’s cultural appeal, accelerates the upgrading of the cultural industry, and advances sustainable urban development [3,4]. Accordingly, enhancing the vitality of historic cultural districts has become a central objective of contemporary urban regeneration and planning.

Scene elements have long served as a key entry point for analyzing urban spatial vitality. Early work by Cervero (1997) [5] examined scene elements from a physical perspective—namely density, mix, and design, the so-called “3D” factors. Ewing (2001) [6] subsequently added destination accessibility and distance to transit, forming a more comprehensive “5D” evaluation system. Since 2016, Daniel (2016) [7] and others have advanced scene theory. They argue that the vitality of historic cultural districts depends not only on physical attributes but also on subjective perception—namely the cognitive and emotional responses formed during cultural experiences in space. Human behavior, therefore, emerges from the interaction between these two dimensions. Previously, scholars also attempted to incorporate subjective perception into district-vitality evaluation frameworks. Zhang et al. [8] assessed how subjective perception, including historical and use values, affects district vitality. Zhang et al. further introduced perceived factors including historical value, authenticity, and cultural experience, highlighting the critical role of subjective perception in shaping spatial vitality [9]. Despite these advances, most existing studies adopt a single-dimensional perspective, focusing either on objective physical factors or on subjective perception, rather than integrating both.

Beyond theoretical limitations, there are also methodological constraints. In recent years, many studies have relied primarily on linear regression approaches, such as Geographically Weighted Regression (GWR), Multiscale Geographically Weighted Regression (MGWR), and Ordinary Least Squares (OLS), to explain the relationship between district vitality and physical factors including location, street texture, and parcel attributes [10]. For example, Jin et al. [11] used GWR and multi-source data from Beijing’s inner city to analyze how the built environment influences multidimensional urban vitality. They found significant associations between vitality and floor area ratio, POI density, and intersection density. Zheng et al. [12] applied OLS and GWR in village-related communities in Nanjing to examine how environmental factors affect public-space vitality for residents and tourists, showing that vitality for both groups is significantly related to accessibility, Green View Index, and leisure facilities. However, these linear models assume relatively stable and additive relationships among variables. As a result, they struggle to detect nonlinear threshold effects, complex interactions, and potential spatiotemporal heterogeneity among scene elements.

Fortunately, in recent years, machine learning models—with strong capacity for nonlinear mappings of feature variables—have been increasingly used to probe the nonlinear threshold effects between historic district vitality and scene elements [13,14,15,16]. Models such as RF, GBDT, LightGBM, and Extreme Gradient Boosting (XGBoost) have attracted growing attention [6,17]. Different models often have distinct advantages in handling nonlinear feature relationships and in explaining feature importance [18]. Through iterative optimization, these models can adaptively adjust feature weights, thereby effectively revealing the threshold effects between spatial vitality and scene elements. Meanwhile, interpretability techniques such as SHAP help clarify each scene element’s contribution and marginal effect, improving the transparency and comprehensibility of the predictions [19].

In sum, this study leverages the nonlinear strengths of machine learning to systematically examine the threshold and synergistic effects of district scene elements on spatial vitality from the dual perspectives of physical factors and subjective perception. By integrating both dimensions into a unified analytical framework, we provide a more comprehensive understanding of district scene characteristics and their complex relationships with vitality. To achieve this goal, we proceed in five steps. First, we construct an explanatory framework that integrates physical and perceptual dimensions of district vitality. We overlay Baidu heatmaps with street-network maps to interpret vitality differences across spatial scales and temporal contexts. Second, we integrate vitality values and scene elements into a unified database for model training. Third, we conduct comparative analysis across multiple machine learning models to identify the optimal model and key scene elements. Fourth, we reveal nonlinear threshold mechanisms and synergistic effects using SHAP-based interpretation. Finally, we propose planning and design strategies to enhance vitality in historic cultural districts.

2. Materials and Methods

2.1. Research Framework

The research framework comprises three parts—data collection and processing, analytical methods and measures, and analytical results and recommendations (Figure 1)—and proceeds in five stages. First, Python (Version 3.1.1) is used to collect Baidu heatmap data from Location-Based Services (LBS), and the ArcGIS (Version 10.8) Raster Calculator is applied to compute and visualize vitality values, producing spatial distribution maps for workdays, rest days, and the combined scenario. Second, leveraging Python, OpenStreetMap (OSM), and Octoparse, we gather Baidu street-view images (BSVIs), Baidu Maps Points of Interest (POIs), district street-network data, and online reviews; these multi-source data are quantified with the semantic-segmentation model SegNet and ArcGIS to extract numeric values of scene elements. Third, scene-element values are integrated with district vitality values into a unified dataset, on which Random Forest (RF), Decision Tree (DT), Gradient-Boosting Decision Tree (GBDT), and Light Gradient Boosting Machine (LightGBM) are trained and compared to determine LightGBM as the optimal model. Fourth, LightGBM is combined with Shapley Additive Explanations (SHAP) to visualize each element’s contribution and to screen high-contribution scene elements. Finally, threshold-effect analysis is conducted on the high-contribution elements to identify critical intervals influencing district vitality, while synergistic-effect analysis reveals element combinations that most effectively stimulate vitality, upon which planning and design recommendations for enhancing vitality in historic cultural districts are proposed.

2.2. Data Sources and Methods

Table 1. Data Sources.

Data Type	Data Source	Data Collection Method
Tianjin Baidu heatmap data	https://lbsyun.baidu.com/, accessed on 25 May 2025	Sampling dates: 25 May 2025 (weekend) and 28 May 2025 (weekday); every two hours from 10:00 to 22:00; 12 heatmap images in total; clear-weather conditions.
POI data for Tianjin central urban districts	https://lbs.amap.com/, accessed on 29 May 2025	District POI data obtained via API, covering commercial, transport, and leisure facilities.
Street-network data for Tianjin central urban districts	https://www.openstreetmap.org/, on 18 May 2025	Street-network layers acquired and subsequently simplified, topologized, and format-standardized.
Street-view imagery for Tianjin central urban districts	https://lbsyun.baidu.com/, accessed on 25 May 2025	Sampling points generated along the street network; BSVIs collected and semantically segmented to extract visual elements and their percentage features.
Online text (review) data	https://www.bazhuayu.com/, accessed on 2 January 2025	User reviews on street experiential perception scraped; time span: May 2022–May 2025.

lists the main data types used in this study, detailed as follows:

(1)

Vitality data. Previous studies have indicated that, under normal conditions, the relative spatial distribution of vitality in historic districts tends to remain stable over short periods. Therefore, selecting one weekday and one weekend day for sampling can, to a certain extent, represent the district’s vitality characteristics under two typical states: routine work-related activities and leisure-oriented activities. Using the Baidu Maps API, we obtained heatmap data for 25 May 2025 (weekend) and 28 May 2025 (weekday). Data collection was conducted under clear weather from 10:00 to 22:00 at two-hour intervals, yielding 12 Baidu heatmap images in total. Following standard processing, we used the ArcGIS Raster Calculator to compute mean vitality values, classified them into seven levels with Jenks natural breaks, and overlaid the results on district maps to produce spatial distributions for weekend, weekday, and comprehensive scenarios.

(2)

Functional-facility data. We acquired Baidu Maps POIs for 2024 via the Baidu Maps API, with preliminary ground-truthing through field surveys.

(3)

Street-network data. Street-network layers were collected from OSM and then merged, simplified, and topologically cleaned.

(4)

Street-view imagery. Sampling points were generated along the street network, and BSVIs were fetched via Python using the Baidu Street View API. We applied the semantic-segmentation model SegNet to segment the BSVIs and extract street visual elements and their percentage features.

(5)

Review data. Using Octoparse and the keywords “Tianjin Wudadao”, “Ancient Culture Street”, and “Italian-Style Street”, we scraped user reviews from Dianping, Ctrip, and WeChat public accounts for the period May 2022 to May 2025.

2.3. Study Area

The study area comprises Tianjin’s Wudadao, Ancient Culture Street, and the Italian-Style District. Using data from China National Knowledge Infrastructure (CNKI) (research platform) and Dianping (consumer platform), we identified these as Tianjin’s research and consumption hot-spot cultural districts; details of data collection and screening are provided in Appendix A, Figure A1 and Figure A2. Inclusion as both a research and consumption hot spot carries dual implications: (i) the district receives priority attention from Tianjin and its culture–tourism authorities, bearing policy-driven tasks of cultural inheritance and renewal; and (ii) the district concentrates consumption and leisure activities, indicating strong visitor attractiveness. Therefore, selecting Wudadao, Ancient Culture Street, and the Italian-Style District not only helps explicate the threshold effects of scene elements on audience attraction but also provides typical cases for revealing the synergy between physical factors and subjective perception and the clustering effects of visitors at different times.

As shown in Figure 2, these districts are key nodes along Tianjin’s Haihe cultural–tourism economic belt. Wudadao covers about 194.6 ha and is characterized by contiguous ensembles of modern Western-style architecture; Ancient Culture Street covers about 20 ha and serves as a major carrier of traditional handicrafts and folk culture; the Italian-Style District covers about 28.45 ha and contains the largest extant concentration of Italianate architecture in China. Using vector boundaries for the three districts, we applied hexagonal grids with cell sizes of 100 m × 100 m (Wudadao), 31 m × 31 m (Ancient Culture Street), and 60 m × 60 m (Italian-Style District), respectively. Because the three districts differ markedly in spatial extent and internal morphology, the grid sizes were adjusted so that the number of valid cells remains within a comparable range across districts (432 spatial units in total after removing null cells), and each cell serves as a fine-grained spatial analysis unit appropriate to the local block scale and functional density. Scene-element metrics were measured within the grid cells to ensure that the unit partitioning better matches the scale of each district. Subsequently, the grid-based observations from the three districts were pooled into a single dataset to estimate an integrated LightGBM model.

2.4. Explanatory Framework for Vitality in Historic Cultural Districts

Drawing on prior studies [7,8,11,20,21,22,23,24,25,26,27,28], we select 11 scene elements—including Street-Network Accessibility, Transport Accessibility, Green View Index, and Facility Mix—to build an explanatory framework for vitality in historic cultural districts (see

Table 2. Explanatory framework combining physical and subjective-perception scene elements for district vitality.

Category	Dimension	Indicator	Indicator Meaning	Measurement Method
Physical elements	Accessibility	Street-Network Accessibility	Ease of walking and internal circulation within the district.	Using ArcGIS, compute the average distance from each district grid cell’s centroid to transport, convenience stores, and other POIs.
		Transport Accessibility	Density/aggregation of public-transport stops (bus/metro) supporting external connection.	Using ArcGIS, compute the mean kernel density of nearby bus- and metro-station POIs for each historic district.
	Visual environment	Green View Index	Proportion of visible greenery in the street view.	Using semantic segmentation of Baidu street-view images, calculate the proportion of vegetation pixels relative to the total number of pixels for each sampling point.
		Spatial Enclosure	Degree of enclosure formed by building façades along streets.	Using semantic segmentation of Baidu street-view images, calculate the proportion of building façade/interface pixels relative to the total number of pixels for each sampling point.
		Sky Openness	Visible-sky proportion in street view.	Using semantic segmentation of Baidu street-view images, calculate the proportion of sky pixels relative to the total number of pixels for each sampling point.
	Functional diversity	Facility Mix	Diversity of facility categories within a grid (land-use/POI mixing).	Using ArcGIS, compute the proportions of each POI category within the historic district.
		Facility Density	Overall concentration/intensity of facilities.	Using ArcGIS, compute the kernel density values for each POI category within the district.
Subjective-perception elements	Facility integration	Traditional–Modern Facility Mix	Share/coordination between traditional and modern formats.	Using ArcGIS, compute the ratio of traditional to modern facilities by dividing the kernel density of traditional-facility POIs (weighted by heritage grade: national = 5, provincial = 3, ordinary = 1) by the kernel density of modern facilities (weighted by rating: 5 = 5, 3 = 3, 1 = 1). The weighting values follow established grading standards reported in the prior literature, and the resulting ratio was further normalized (Min–Max) to improve comparability and mitigate potential subjectivity in indicator construction.
	Cultural expression	Heritage Attractiveness	Perceived attention/knowledge regarding historic remains.	Using ArcGIS, compute the kernel density of heritage-related POIs within each grid cell and multiply it by the visitor-attention weight for the heritage attractiveness dimension, derived from the proportion of heritage attractiveness-related high-frequency words in the overall review word-frequency statistics.
	Cultural display	Exhibition-Facility Attractiveness	Perceived attention to exhibition/display facilities.	Using ArcGIS, compute the kernel density of exhibition-facility-related POIs within each grid cell, and multiply it by the visitor-attention weight for the Exhibition-Facility Attractiveness dimension, derived from the proportion of Exhibition-Facility Attractiveness-related high-frequency words in the overall review word-frequency statistics.
	Cultural leisure	Leisure-Facility Attractiveness	Emotional engagement with leisure facilities and activities.	Using ArcGIS, compute the kernel density of Leisure-Facility Attractiveness-related POIs within each grid cell and multiply it by the visitor-attention weight for the Leisure-Facility Attractiveness dimension, derived from the proportion of Leisure-Facility Attractiveness-related high-frequency words in the overall review word-frequency statistics.

). To enhance theoretical coherence and reproducibility, these indicators are organized into two overarching dimensions derived from scene theory: (i) physical elements, reflecting objective spatial structure and functional configuration; and (ii) subjective-perception elements, reflecting visitors’ cognitive evaluation and experiential responses. The selection of indicators is based on three criteria: theoretical relevance to vitality formation, empirical support in the prior literature, and measurability at the hexagonal-grid scale. On the physical side, given that historic cultural districts typically cover street segments or small precincts with relatively clear internal boundaries and functional profiles, we adopt three dimensions to capture form and structure: accessibility, visual environment, and functional diversity. On the subjective-perception side, we innovatively incorporate scene-related evaluative indicators—authenticity, theatricality, and legitimacy—and, following localized applications in the literature, operationalize them through four dimensions, facility integration, cultural expression, cultural display, and cultural leisure, reflecting how visitors perceive cultural value and experiential ambience. In this study, these subjective-perception elements are operationalized by combining the spatial distribution of different types of facilities with visitor-attention weights extracted from review texts so that they reflect aggregated patterns of what visitors emphasize in relation to heritage, exhibition, and leisure experiences in the three historic cultural districts. After quantitatively deriving the scene elements as described below, we compile samples at the level of hexagonal grid cells, ensuring that the sampling extent matches the scale of each district.

2.5. Machine-Learning Models

2.5.1. Model Training

Evidence shows that, for small- to medium-sample studies in spatial socioeconomics, ensemble learning models outperform traditional linear methods [29,30]. We split the sample into training and test sets in an 8:2 ratio and trained four models: RF, DT, GBDT, and LightGBM. Guided by sample size and prior research [18], we set the main hyperparameters as follows: learning rate {0.005, 0.01, 0.02, 0.05}; maximum tree depth 1–40. We then applied random search with five-fold cross-validation to determine the optimal hyperparameters for each model [31], aiming to control model complexity and prevent overfitting. Model selection relied on comparative evaluation metrics (

Table 3. Optimal hyperparameter settings and test-set evaluation results for different decision-tree models.

Model Type	Learning Rate	Max Depth	MAE	RMSE	R2
RF	0.05	10	0.0008	0.0012	0.72
DT	0.05	17	0.0009	0.0014	0.62
GDBT	0.05	6	0.0008	0.0011	0.74
LightGBM	0.05	16	0.0005	0.0010	0.85

). Higher R² indicates better fit, while lower MAE and RMSE indicate better predictive performance [32]. LightGBM achieved the best overall results (MAE = 0.0005, RMSE = 0.0010, R² = 0.85), outperforming RF (R² = 0.72), DT (R² = 0.62), and GBDT (R² = 0.74). To further assess model robustness and mitigate potential bias arising from a single random split, we conducted repeated five-fold cross-validation (10 repetitions) for the three analytical scenarios. The results demonstrate stable predictive performance across different data partitions: weekend (mean R² = 0.888, SD = 0.027, range: 0.822–0.933), comprehensive (mean R² = 0.877, SD = 0.028, range: 0.813–0.924), and weekday (mean R² = 0.857, SD = 0.031, range: 0.787–0.911). The relatively small standard deviations and narrow performance ranges indicate that the predictive superiority of LightGBM is robust and not driven by a particular data split. Therefore, we use LightGBM as the primary model to estimate the threshold and synergistic effects of district scene elements on district vitality.

2.5.2. LightGBM Model

LightGBM has been widely used in urban-data modeling and prediction in recent years [33]. It is built on the GBDT framework and incorporates a histogram-based efficient algorithm together with a leaf-wise splitting strategy [34], which accelerates training and reduces memory consumption. To prevent overfitting, LightGBM limits the maximum tree depth and performs feature splits by selecting the leaf node with the largest gain, growing the tree layer by layer. During training, the model iteratively adds new weak learners (single decision trees) to fit the residuals from the previous round, thereby progressively approaching the true labels. Suppose that in the m-th iteration the objective function is as follows:

$$L\left(y_i,{\widehat{y}}_i\right)$$

(1)

The corresponding prediction update formula is as follows:

$${{\widehat{y}}_i}^{\left(m\right)}={{\widehat{y}}_i}^{(m-1)}+\eta f_m\left(x_i\right)$$

(2)

The overall objective of LightGBM typically consists of a loss term and a regularization term and can be written as follows:

$${Objective}^{\left(m\right)}=\sum _{i=1}^{N}L\left(y_i,{{\widehat{y}}_i}^{\left(m\right)}\right)+\sum _{m=1}^M\Omega \left(f_m\right)$$

(3)

where L is the loss function measuring prediction error, and Ω is the regularization term used to control model complexity. The final model output is the weighted sum of all weak learners:

$${\widehat{y}}_i=\sum _{m=1}^M\eta f_m\left(x_i\right)$$

(4)

In this study, LightGBM’s high-dimensional modeling and feature-selection capabilities help capture the complex nonlinear relationships through which scene elements affect district vitality. Compared with deep neural network architectures that often prioritize domain-specific predictive performance—such as recursive long short-term memory models for nonlinear structural seismic response prediction and DSYOLO-style segmentation networks for tunnel lining crack detection [35,36]—Gradient-Boosting Decision Trees offer a favorable balance between accuracy and interpretability, which is crucial for explaining how scene elements jointly shape vitality patterns in historic cultural districts.

2.5.3. SHAP

As a game-theoretic, post hoc interpretability method, SHAP is commonly coupled with machine learning models to address the credibility issues of complex “black-box” predictors. To enhance interpretability, we apply SHAP to provide a posterior explanation of the LightGBM results [37]. SHAP delivers global interpretability via summary plots, enabling a comprehensive understanding of feature importance across all predictions [38]. For example, the summary plot presents the average magnitude and direction of each feature’s influence. In addition, by quantifying each input feature’s contribution to an individual prediction, SHAP offers local interpretability [39]. Accordingly, we use SHAP visual tools—including feature-importance plots, dependence plots, and interaction plots—to interpret the threshold and synergistic effects of scene elements on district vitality. SHAP values, based on Shapley values, are defined as follows:

$${\varphi }_i\left(f,x\right)=\sum _{S\subseteq F\backslash \left\{i\right\}}{\displaystyle \frac{\left|S\right|!\left(\left|F\right|-\left|S\right|-1\right)!}{\left|F\right|!}}\left[f_x\left(S\cup \left\{i\right\}\right)-f_x\left(S\right)\right]$$

(5)

where ϕ_i(f, x) denotes the marginal contribution of feature i to the prediction for sample x, F is the full set of features, S is a subset that does not include feature i, and f_x(S) is the model’s prediction given subset S. This formulation computes the average influence of feature i across all coalitions and is the only attribution method that satisfies efficiency and fairness.

3. Results

3.1. Spatial Vitality of Historic Cultural Districts

To capture temporal differences in the spatial distribution of district vitality, we visualize Baidu heatmap data separately for weekday, weekend, and comprehensive scenarios. Using Jenks natural breaks, vitality is classified into five categories by heatmap intensity: high vitality, relatively high vitality, moderate vitality, relatively low vitality, and low vitality. Figure 3, Figure 4 and Figure 5 present the spatial distributions for the three scenarios, where the comprehensive map uses the average of weekend and weekday vitality.

In Wudadao, the contrast between high- and low-vitality areas is pronounced. Under the comprehensive and weekend scenarios, high values cluster along the Munan Park—Minyuan Square—Chongqing Road corridor, where dense historic architecture, cultural–tourism landmarks, cafés, and cultural–creative shops form a continuous high-heat belt. Minyuan Square functions as the core node for holiday footfall and consumption. By contrast, the southern and southwestern edges—such as Machang Road and Kunming Road—remain persistently low in vitality, dominated by residential and educational uses, lacking cultural–tourism guidance and active street interfaces, which creates a clear vitality fracture. On weekdays, high vitality contracts further, remaining only on primary streets and core nodes, while edge areas are broadly low in vitality, indicating imbalanced spatial use.

In Ancient Culture Street, spatial vitality follows a pattern of “higher in the north, lower in the south; concentrated core, attenuated edges”. High vitality concentrates at the northern gateway adjacent to major urban roads, where Transport Accessibility is strong and traditional-culture facilities cluster, drawing heavy tourist flows. Riverside nodes such as Jinmen Opera House also attract crowds by leveraging landscape resources. The central zone maintains moderate vitality through shop and workshop agglomeration, while the Tianhougong Temple and its southern surroundings remain relatively low. Temporally, weekday vitality is concentrated in the north and along riverside nodes, showing core dependency; on weekends, vitality expands toward the center and south as cultural consumption and tourism demand are released, markedly lifting overall heat levels.

In the Italian-Style District, the pattern is “active rings with a central trough”, with a sharp contrast between high and low vitality. In the comprehensive and weekend maps, high vitality is concentrated around Guangfu Road, Ziyou Road, and Minzu Road, where nodes such as the Tianjin Urban Planning Exhibition Hall and Marco Polo Square, together with surrounding commercial facilities, form a relatively continuous high-vitality ring. Especially on weekends, cultural–tourism consumption and leisure activities increase significantly, and Minyuan Street and Ziyou Road develop into holiday crowd corridors. By contrast, the northern sector—such as Guangming Road and Minzhu Road, with Italian celebrity residences and housing—shows persistently low vitality, characterized by enclosed buildings and low-frequency spaces lacking open interfaces and supporting formats, thus forming evident vitality depressions. On weekdays, vitality weakens further, with high-vitality areas shrinking to the cultural–tourism core nodes and most surrounding tracts performing poorly.

3.2. Relative Importance of Scene Elements for District Spatial Vitality

To reflect temporal differences in how scene elements affect district vitality, we trained separate datasets for the weekday, weekend, and comprehensive scenarios and obtained comparable model results. Using SHAP, we ranked the relative importance of all scene elements. As shown in Figure 6, Figure 7 and Figure 8, blue bars indicate each element’s average contribution to spatial vitality, and longer bars denote higher importance. Each scatter point represents one sample; its horizontal position is the SHAP value, and the color encodes the element’s magnitude with a blue-to-red gradient, where higher values are closer to red. Physical elements contribute substantially to district vitality (mean RI = 62.15%). Among them, Street-Network Accessibility consistently dominates across scenarios (mean RI = 51.95%). This accords with prior research showing that higher Street-Network Accessibility improves spatial enterability and service convenience, thereby reinforcing crowd aggregation and spatial vitality [38]. Subjective-perception elements also exert strong influence (mean RI = 37.85%). Notably, Leisure-Facility Attractiveness and Exhibition-Facility Attractiveness remain steadily ranked second and third across time periods, with an average combined contribution of 33.98% to district vitality.

We conduct focused analyses of the top six elements by relative importance (i.e., SHAP values) for each scenario (Figure 9, Figure 10 and Figure 11). For weekend vitality, the key elements are Street-Network Accessibility, Leisure-Facility Attractiveness, Exhibition-Facility Attractiveness, Transport Accessibility, Traditional–Modern Facility Mix, and Facility Density. For weekday vitality, the key elements are Street-Network Accessibility, Leisure-Facility Attractiveness, Exhibition-Facility Attractiveness, Facility Density, Transport Accessibility, and Traditional–Modern Facility Mix. For comprehensive vitality, the key elements are Street-Network Accessibility, Leisure-Facility Attractiveness, Exhibition-Facility Attractiveness, Traditional–Modern Facility Mix, Transport Accessibility, and Spatial Enclosure.

On the physical elements side, Street-Network Accessibility shows a strong positive impact across all time periods (comprehensive RI = 51.6%, weekday RI = 48.50%, weekend RI = 44.40%), with the largest effect on the comprehensive scenario. Transport Accessibility has a more pronounced positive effect on weekends (RI = 5.9%), while its role is relatively weaker in the comprehensive (RI = 3.70%) and weekday (RI = 4.90%) scenarios. Facility Density is positively associated with vitality on both weekdays (RI = 7.90%) and weekends (RI = 3.30%), with a stronger effect on weekdays. Spatial Enclosure exhibits a slight negative association in the comprehensive scenario (RI = 2.50%), which may indicate a preference for leisure activities in more open spaces, where enclosed spaces are less conducive to vitality gains. On the subjective-perception elements side, Leisure-Facility Attractiveness ranks second across all time periods (comprehensive RI = 21.80%, weekday RI = 19.80%, weekend RI = 27.90%), indicating a consistent and stable positive influence, with the strongest contribution on weekends (RI = 27.90%). Exhibition-Facility Attractiveness ranks third across scenarios (comprehensive RI = 8.70%, weekday RI = 8.90%, weekend RI = 7.40%), showing a consistently positive effect and a slight uptick on weekdays. Traditional–Modern Facility Mix has a positive impact on weekend vitality (RI = 4.00%) but exerts relatively weaker effects on weekdays (RI = 2.70%) and in the comprehensive scenario (RI = 2.50%).

3.3. Threshold Effects of Scene Elements

SHAP dependence plots (Figure 12, Figure 13 and Figure 14) reveal the nonlinear impacts of scene elements on spatial vitality. The horizontal axis shows each element’s value, and the vertical axis shows its contribution to overall vitality. Larger values on the vertical axis indicate stronger positive effects.

We conduct threshold analysis for the top six key elements, which together account for more than 80 percent of total contribution. On the side of physical elements, Street-Network Accessibility is strongly and positively associated with vitality across all time periods. When its value is below 0.3, contributions are generally weak. As it approaches 0.4, vitality rises rapidly in all districts and scenarios. The weekend scenario shows the largest marginal gains, reaching a vitality value of about 0.004 when accessibility is 0.7. This indicates that Street-Network Accessibility is a core element at all times and is especially beneficial for weekend vitality. Transport Accessibility exhibits a U-shaped nonlinear pattern across time periods. When the density value is below 0.3, vitality responds negatively; once it exceeds 0.5, the contribution turns positive and then stabilizes, with the strongest rebound on weekdays. Facility Density shows a kinked positive growth on both weekdays and weekends. Below 0.3, contributions are limited; once this threshold is crossed, vitality increases rapidly, peaking at about 0.0003 on weekdays and slightly lower at about 0.00015 on weekends. This suggests higher sensitivity of space use efficiency and vitality generation to facility agglomeration on weekdays. Spatial Enclosure has a slight effect on comprehensive vitality. Contributions peak when enclosure lies between 0.2 and 0.3, then decline, with some samples turning negative.

On the side of subjective-perception elements, Traditional–Modern Facility Mix shows a mid-value peak across time, with the strongest marginal effects when the mix lies between 0.6 and 0.8. The peak is highest in the comprehensive scenario, while weekend contributions are relatively stable with smaller fluctuations. Leisure-Facility Attractiveness displays a rapid rise followed by a plateau in all time periods. When its value is between 0 and 0.2, vitality increases quickly across districts. The weekend response is the largest, reaching about 0.0015 when the value is 0.2. Weekday and comprehensive responses are weaker at about 0.0006 and 0.0007, respectively. After the value exceeds 0.3, marginal effects level off, and differences across scenarios diminish. Exhibition-Facility Attractiveness exhibits an inverted U-shaped pattern over time. Vitality peaks when the value ranges from 0.05 to 0.15. Once it exceeds 0.20, the effect turns negative, leading to a decline in vitality.

3.4. Interaction Effects Among Scene Elements

By computing interaction values for pairs of scene elements, this study illustrates how elements jointly influence district spatial vitality and presents the results as interaction matrices (Figure 15, Figure 16 and Figure 17). The matrices display interactions among the top six elements by contribution share. In each figure, the left column lists the selected variable, the top row lists the interacting variable, the x-axis shows the Shapley interaction value, red indicates a positive (promoting) effect, and blue indicates a negative (suppressing) effect; larger values are closer to red.

To identify combinations of scene elements that most efficiently enhance spatial vitality, we selected the top three element pairs by SHAP interaction values under the weekend, weekday, and comprehensive scenarios and plotted their interaction effects (Figure 18, Figure 19 and Figure 20). The interaction plots illustrate how variables jointly influence urban vitality. On each plot, the X-axis shows the value of the primary variable, the color gradient represents changes in the interacting variable, and the Y-axis indicates the direction and magnitude of their combined contribution to spatial vitality.

On weekends (Figure 18), when Leisure-Facility Attractiveness >0.35 and Street-Network Accessibility lies within 0.5–0.8, their combination markedly promotes district vitality; conversely, when Street-Network Accessibility <0.2 or Leisure-Facility Attractiveness <0.15, vitality is suppressed. At the same time, moderate Transport Accessibility (0.2–0.4) amplifies the positive effect of Leisure-Facility Attractiveness, whereas very low accessibility (<0.1) reduces vitality. In addition, when Traditional–Modern Facility Mix >0.6 and Leisure-Facility Attractiveness >0.2, vitality increases significantly, indicating that cultural plurality elevates the marginal contribution of leisure elements. On weekdays (Figure 19), the largest gains occur when Leisure-Facility Attractiveness >0.2 and Transport Accessibility remains within 0.2–0.35. If Transport Accessibility is too low (<0.1), even high Leisure-Facility Attractiveness has difficulty generating vitality. The coupling of Leisure-Facility Attractiveness and Exhibition-Facility Attractiveness shows that vitality grows markedly when Leisure-Facility Attractiveness >0.25 and Exhibition-Facility Attractiveness >0.1, whereas insufficient Leisure-Facility Attractiveness (<0.15) weakens the utility of Exhibition-Facility Attractiveness. The interaction between Street-Network Accessibility (>0.6) and Leisure-Facility Attractiveness (>0.3) is comparatively mild on weekdays. Under the comprehensive scenario (Figure 20), vitality rises steadily when Leisure-Facility Attractiveness >0.3 and Exhibition-Facility Attractiveness >0.15; when Leisure-Facility Attractiveness <0.15, Exhibition-Facility Attractiveness cannot maintain sustained crowd aggregation. The interaction between Street-Network Accessibility and Leisure-Facility Attractiveness mirrors the weekend pattern but with a slightly smaller increase in vitality. Finally, when Transport Accessibility lies within 0.2–0.35 and Leisure-Facility Attractiveness >0.35, vitality remains stable; if Transport Accessibility is very low, vitality is suppressed even when Leisure-Facility Attractiveness is high.

4. Discussion

4.1. Thresholds and Synergies of Scene Elements for District Spatial Vitality

4.1.1. Threshold Effects of Individual Scene Elements

This study finds pronounced threshold effects for key elements such as Street-Network Accessibility, Facility Density, and Leisure-Facility Attractiveness in shaping spatial vitality, consistent with prior evidence that the built environment exerts nonlinear influences on district vitality [18,40]. We further reveal the temporal conversion of these thresholds across weekday–weekend–comprehensive scenarios, extending existing explanations of the time dependence of vitality formation.

Street-Network Accessibility functions as a core condition across all periods, with a particularly strong role on weekends, which accords with earlier findings [41]. This higher weekend contribution may be attributed to its role in facilitating internal pedestrian circulation and slow traffic movement within historic districts. Compared with Transport Accessibility—which primarily reflects external connectivity to the broader urban system—Street-Network Accessibility captures the structural permeability and walkable continuity of the internal street fabric. During weekends, when leisure-oriented and exploratory activities increase, the quality of internal circulation becomes more critical for sustaining spatial vitality. Unlike studies that focus primarily on the spatial utility of accessibility, we emphasize its marginal advantage under different temporal scenarios, indicating that on weekends people’s leisure and consumption behaviors depend more strongly on convenient access to facilities. Transport Accessibility shows marked time-of-day differences and tends to stabilize beyond a certain level, with the strongest rebound effect on weekdays, consistent with empirical results in the literature [42]. Our analysis further indicates that when Transport Accessibility is either too high or too low, its effect on vitality may turn negative, suggesting that cultural and leisure activities require an appropriate level of footfall rather than mere transport concentration. Spatial Enclosure promotes vitality within a moderate range but can become detrimental when it exceeds that range. This result supports Gehl’s (2011) notion of a “perceivable boundary”, implying that moderate enclosure fosters safety and comfort, whereas excessive enclosure constrains openness and interaction [43]. Facility Density exhibits a kinked positive trend on both weekdays and weekends and is more sensitive on weekdays, aligning with the view that facility agglomeration enhances service availability and thereby promotes vitality [10]. We also find that weekday sensitivity is higher, indicating that residents’ everyday activities rely more on efficient facility supply, while weekend vitality is more strongly supported by Leisure-Facility Attractiveness. Traditional–Modern Facility Mix significantly enhances vitality within a moderate range but becomes counterproductive when excessively high. This suggests that overcommercialization and homogenization may erode the uniqueness of historic cultural districts, a concern echoed by prior studies on cultural loss under commercial development [44]. When Leisure-Facility Attractiveness reaches a certain threshold, vitality rises rapidly across time periods, with the strongest response on weekends. This aligns with research suggesting that leisure experiences substantially strengthen visitors’ perception of cultural value [45]. The temporal variation in its contribution may reflect differences in behavioral motivations: on weekdays, spatial activities are often structured around commuting and work-related routines, resulting in more purpose-driven mobility patterns, whereas on weekends, leisure, social interaction, and experiential consumption become primary drivers of urban activity, making the attractiveness of leisure facilities more influential in shaping vitality. However, when Leisure-Facility Attractiveness exceeds 0.3, vitality gains taper and inter-period differences narrow, indicating saturation. Exhibition-Facility Attractiveness has a positive effect within 0.05–0.15 but shows limited gains when attraction becomes excessive. Field observations suggest that historic cultural districts depend more on diversified and interactive display modes. Overly one-way exhibition formats or closed display spaces—for example, clusters of closed celebrity residences in Wudadao and the Italian-Style District—are not conducive to enhancing spatial vitality.

4.1.2. Synergistic Effects of Scene Element Combinations

This study shows that the interactions among Leisure-Facility Attractiveness, Street-Network Accessibility, and Transport Accessibility exert significant influence on vitality in historic cultural districts. When accessibility combines with cultural elements, the vitality effect is amplified, and this is especially evident in tourism and leisure contexts [46]. Our results confirm this relationship. They further indicate that the interaction between Leisure-Facility Attractiveness and Street-Network Accessibility not only strengthens vitality but also exhibits clear temporal variation. On weekends, this interaction becomes more pronounced. This suggests a stronger dual dependence on convenience and cultural ambience in leisure-consumption settings. Prior work shows that Exhibition-Facility Attractiveness enhances district appeal [47]. Our findings are consistent and also reveal that when Leisure-Facility Attractiveness is insufficient, the effect of Exhibition-Facility Attractiveness weakens. This suggests that exhibition attraction alone cannot sustain crowds and must be coupled with leisure-consumption scenes. When Transport Accessibility lies between 0.2 and 0.4, it markedly amplifies the positive effect of Leisure-Facility Attractiveness, promoting vitality generation. This aligns with evidence that Transport Accessibility correlates strongly with district vitality [42]. Beyond confirming this correlation, our study uncovers a nonlinear interaction between Transport Accessibility and Leisure-Facility Attractiveness. We also find that vitality increases significantly when districts combine Traditional–Modern Facility Mix with leisure activities. This supports the view that cultural diversity enhances overall district appeal [48]. Under a plural cultural background, leisure consumption not only improves experiential quality but also strengthens the sustainability of vitality. Under the comprehensive scenario, interactions among elements become milder yet remain positive. This suggests that long-term vitality depends more on the joint action of cultural and spatial elements than on the improvement of any single factor.

It should be noted that these interaction effects are derived from a single LightGBM model estimated on the pooled sample of all three historic cultural districts. In other words, they describe interaction relationships among scene elements as identified from a unified model fitted to the combined grid-based observations of Wudadao, Ancient Culture Street, and the Italian-Style District, rather than effects estimated separately for each district. The district-specific vitality heatmaps in Figure 5, however, visually indicate that high-vitality cells are organized in different spatial configurations in the three districts: for example, Wudadao tends to form more continuous high-vitality corridors along major streets, whereas Ancient Culture Street and the Italian-Style District show more compact clusters around core cultural and commercial nodes, accompanied by distinct combinations of Spatial Enclosure, sky openness, and cultural/leisure facilities. These observations suggest that the general interaction relationships identified by the pooled model may be modulated by local spatial morphology and cultural functions.

4.2. Optimization Recommendations and Insights for Scene Elements Driving District Spatial Vitality

This study emphasizes optimizing scene elements from both the physical elements and subjective-perception elements dimensions to more effectively enhance spatial vitality in historic cultural districts. To address the marginalization of vitality in areas with insufficient Facility Density or Street-Network Accessibility, small-scale walkable amenities should be introduced. In the southern segment of Ancient Culture Street and along the western edge of Wudadao, mobile coffee carts, pop-up book bars, and temporary stalls could be deployed to create a 300 m “service walking circle”. This would ensure that residents and visitors can access food, rest areas, and drinking water within five minutes. Streetfront shops are encouraged to share space resources, extend opening hours during off-peak times, and add temporary resting points. The U-shaped effect of Transport Accessibility suggests differentiated intervention strategies. At low accessibility levels, vitality tends to be suppressed; at higher levels, crowd aggregation increases. Therefore, micro-scale interventions should aim to improve travel experience rather than simply increase transport intensity. Shared-bike docks and simple waiting areas can be placed at bus stops and nodal plazas to increase transfer efficiency. Wayfinding signs and QR-code plaques may be added along main pedestrian routes to shorten visitors’ learning time of spatial layouts. In segments with high Transport Accessibility, such as key areas of Wudadao, temporary pedestrian corridors can be opened during holidays to reduce congestion and enhance walkability. Within residential living circles, convenience stores, fast-food outlets, and resting points should be arranged to meet daily needs. During holidays, markets or cultural–creative fairs may be organized in core plazas near neighborhoods to diversify local cultural consumption and activate underused spaces. Findings for Spatial Enclosure indicate that values between 0.20 and 0.30 best enhance feelings of safety and willingness to linger, whereas excessive enclosure above 0.35 suppresses vitality. Accordingly, squares and nodal areas should use greenery, seating, and paving design to create moderate enclosure. In alleys and subareas—such as parts of western Wudadao—sections of enclosed interfaces may be partially opened. A combination of semi-enclosed and open interfaces can help balance safety and openness. To prevent homogenization driven by an overly high Traditional–Modern Facility Mix, the share of standardized chain stores should be carefully controlled. At the same time, uses closely tied to local culture—such as intangible-heritage workshops and independent craft studios—should be encouraged. An entry and approval mechanism can evaluate cultural fit for new tenants, and festivals and creative fairs can highlight a “tradition plus innovation” model by combining traditional crafts with interactive installations to avoid monotone cultural expression. Finally, Leisure-Facility Attractiveness and Exhibition-Facility Attractiveness should be strengthened through interactive and experiential modes rather than static display alone. In cafés and bookstores in Wudadao and the Italian-Style District, small performances and hands-on craft sessions can foster participation in cultural consumption. In areas with dense exhibition facilities, guided interpretation, immersive shows, and art workshops can raise engagement. Ancient Culture Street and the Italian-Style District can further utilize historic courtyards and small street-corner spaces to host periodic immersive performances, street-art shows, and mini concerts. These activities would enhance interactivity and maximize the generative effects of spatial vitality.

4.3. Limitations and Future Directions

This study has several limitations that warrant improvement in future work. On data and methods, we rely primarily on POI data to characterize facilities and cultural elements. This approach cannot fully reflect actual usage frequency or the dynamic behaviors of people, which may lead to overestimation or underestimation of vitality in some areas. Moreover, the subjective cognitive indicators derived from online-review word-frequency weighting represent an indirect proxy of human perception; future work could incorporate questionnaire-based calibration or controlled perception surveys to further strengthen construct validity. In addition, because the samples are organized as adjacent hexagonal grids, there is inherent spatial dependence and the observations are not strictly independent, so the reported performance metrics should not be interpreted as strict out-of-sample prediction accuracy but rather as supporting the use of the LightGBM–SHAP framework as an explanatory tool to reveal the relative importance, nonlinear effects, and interactions of scene elements. On study scope, the analysis focuses on typical historic cultural districts in Tianjin’s central city, without cross-city or cross-type comparisons, so the generalizability of the conclusions remains limited. In light of these shortcomings, future research can advance in three directions. First, incorporate more diverse and dynamic data sources, such as mobile signaling and time-series trajectories, to depict human behavior and cultural experience more comprehensively. Second, conduct comparative analyses across scales and across cities to reveal both common patterns and differentiating mechanisms of vitality generation under different spatial contexts, thereby offering more generalizable references for the renewal and governance of historic cultural districts. Third, adopt spatial-block cross-validation schemes or spatially explicit modeling approaches that explicitly account for spatial autocorrelation among neighboring grids, so as to further test the robustness and transferability of the modeling results.

5. Conclusions

This study examines the spatial heterogeneity of vitality and its underlying nonlinear mechanisms in hot-spot historic cultural districts in central Tianjin from the dual perspectives of physical elements and subjective-perception elements. The results reveal clear temporal and spatial differences in vitality and identify threshold effects and synergistic effects among scene elements. Key determinants include Street-Network Accessibility, Leisure-Facility Attractiveness, Exhibition-Facility Attractiveness, Transport Accessibility, Traditional–Modern Facility Mix, Facility Density, and Spatial Enclosure, which together account for about 80 percent of total contribution.

On the physical side, Street-Network Accessibility is a core driver across time and shows stronger marginal effects on crowd aggregation and lingering on weekends. Transport Accessibility follows a U-shaped pattern in which very low density cannot sustain aggregation, while a moderate range markedly amplifies the positive roles of leisure and exhibition functions. Facility Density is more sensitive on weekdays, indicating stronger dependence of daily commuting and consumption on efficient facility supply.

On the subjective-perception side, Leisure-Facility Attractiveness contributes little at low levels, but once it passes a critical value, vitality rises rapidly with the strongest response on weekends. Traditional–Modern Facility Mix in the 0.6 to 0.8 range enhances distinctiveness and vitality, whereas excessively high levels may indicate overcommercialization and homogenization that weaken overall appeal. The analysis further shows that interactions between physical elements and subjective-perception elements amplify positive effects on vitality. For example, the interaction between Street-Network Accessibility and Leisure-Facility Attractiveness generates contributions that exceed those of single elements. These interaction effects are generally conditioned by thresholds. Only when relevant elements exceed critical points do they strengthen vitality, while values below thresholds may turn the effect negative. For instance, the interaction between Transport Accessibility and Exhibition-Facility Attractiveness produces synergy within a moderate range, yet when Transport Accessibility is very low, even high Exhibition-Facility Attractiveness struggles to sustain vitality.

Taken together, the findings contribute to the literature in two respects. First, they establish a dual-dimensional explanatory framework that integrates objective spatial structure and subjective experiential evaluation within a unified vitality model. Second, they provide empirical evidence that vitality formation in historic cultural districts is not linear but governed by threshold-sensitive and interaction-based mechanisms.

Overall, the findings underscore that spatial vitality in historic cultural districts depends on the synergy between physical elements and subjective-perception elements and that vitality is generated through time-dependent, threshold-based, and interaction-driven mechanisms. This study deepens understanding of how the configuration of scene elements shapes vitality and provides empirical support for activating cultural resources and enhancing district vitality. It is intended to inform planning practice so that material space and cultural resources evolve in concert, highlight local characteristics, and advance the sustainable development of historic cultural districts.