Analysis and Optimization of the Spatial Patterns of Commercial Service Facilities Based on Multisource Spatiotemporal Data and Graph Neural Networks: A Case Study of Beijing, China

Abstract

As a crucial component of urban economic activities, the layout and optimization of urban commercial spaces directly influence the economic prosperity and quality of life of residents. Therefore, comprehensively and accurately characterizing the distribution characteristics and evolutionary patterns of urban commercial spaces is essential for improving the efficiency of urban spatial allocation and achieving scientific spatial planning and governance. This paper utilizes multisource spatiotemporal data, employing geographic spatial analysis methods and graph neural network models to explore the spatial structure of commercial service facilities in Beijing and their relationships with population density and land use, thereby achieving a detailed classification of the commercial service patterns at the natural neighborhood scale. The research findings indicate a significant association between commercial service facilities and population, as well as land use, with a strong spatial heterogeneity. There exists a dissonance between the layout of commercial service facilities and population distribution, and the differences in commercial service development across various regions pose challenges to balanced urban development. Based on this, this paper provides specific recommendations for optimizing the urban commercial spatial structure, offering reference points for future urban planning and development.

1. Introduction

The urban space is composed of various sub-functional spaces and serves as one of the most critical places for human activities [1]. As the core of metropolitan functions, with the rapid development of commerce and services, urban commercial districts have become the most dynamic spaces within cities [2,3]. The role of commercial services in urban spaces has evolved from merely being auxiliary facilities to becoming a primary driver of urban economic development. They now stimulate the growth of other infrastructure and significantly influence the city’s economy, society, and culture [4].

Commercial facilities play a crucial role in economic growth and urban development in cities [5]. The well-planned layout of commercial facilities contributes to the city’s economic development and directly impacts the quality of life for urban residents. Understanding the spatial distribution patterns of urban commercial facilities enables the rational allocation of urban resources and commercial infrastructure, thereby promoting healthy economic growth in the city [6,7].

As a core topic in urban geography, the spatial patterns and dynamic changes in urban commercial spaces have long been a focus of academic attention. In recent years, with the rapid development of big data and spatial analysis technologies, numerous studies have delved into the spatial layout of urban commercial facilities and the driving factors behind these patterns [8,9]. The spatial distribution of urban commercial outlets is influenced by multiple factors, including population, economic conditions [10], urban road networks [11], public transportation [9], and other geographical factors [12]. For example, Zhou et al. [13] used the co-location quotient to measure the spatial association characteristics and patterns between commercial land and residential areas in Beijing, revealing the spatial asymmetry of the mutual attraction between residential and commercial activities. Qu et al. [14] used geo-detectors to reveal that different leisure tourism modes exhibit spatial scale dependence. Similarly, the spatial distribution of commercial outlets also influences urban population mobility, thereby impacting the vitality of the city [5]. Therefore, an essential challenge for urban geographers and city planners is how to adapt the urban commercial structure to the socio-economic transformation while balancing the needs of urban residents and promoting the healthy and orderly development of cities.

Studying the structure of urban commercial spaces requires identifying and classifying commercial districts. Point of interest (POI) data, as a type of point data rich in geospatial information, represents the flow and clustering of various urban elements and the spatial distribution of urban functions through its characteristics and aggregation patterns. Additionally, it offers the advantages of easy accessibility and processing. With the advancement of big data acquisition and analysis technologies, POI data have become widely utilized in research on urban spatial structures [14,15,16]. By leveraging POI data in conjunction with other multisource geospatial data, a new perspective is provided for quantifying the spatial layout of urban commercial spaces and their spatial relationships with other geographic elements. Wang et al. [4] proposed a method for identifying and classifying urban commercial districts based on POI and road network data. Using the Huff model and Voronoi method, they analyzed the spatial distribution patterns of commercial districts according to population spatial patterns, offering recommendations for optimizing the structure of urban commercial spaces. Zhou et al. [17] used commercial and residential POI data to quantify the spatial association between them, providing development suggestions for renovating urban commercial–residential spaces and constructing human-centered smart cities. Shi et al. [11] combined road network data and commercial POI data, using measures such as network centrality, kernel density estimation, and Pearson correlation coefficients, to explore the relationship between road network centrality and the spatial layout of different types of commercial facilities. Zhang et al. [18], using multi-source POI data from residents, identified changes in the relationship between residential and commercial facilities in shrinking cities from a supply–demand perspective, driven by population loss and economic decline.

In recent years, the application of machine learning in urban geography has experienced rapid growth. Urban geography primarily focuses on the study of urban spatial structure [19], economic activities [20], and transportation networks [21]. With its powerful data processing and pattern recognition capabilities, machine learning brings new breakthroughs to these research areas. By integrating multi-source data such as remote sensing imagery [22], street-level imagery [23], and social media data [24,25], machine learning algorithms can automatically extract and analyze characteristics of the urban environment, including land use classification [26], building function classification [27], identification of urban functional zones [28,29,30], and vehicle speed and flow estimation based on street-view imagery [31]. However, effectively integrating the theoretical framework of urban geography with the technical advantages of machine learning remains a significant challenge in current research. Previous studies on the spatial patterns of commercial facilities had not sufficiently considered the spatial dependencies and interactions between blocks. When evaluating the commercial status of a neighborhood, the influence of surrounding neighborhoods on the target neighborhood was often overlooked. Additionally, population density data had typically been treated as a static feature, neglecting the significance of temporal variations in population density when assessing the spatial distribution of commercial services. Since the units are not isolated, correlations exist between adjacent units. Therefore, this study employs the Graph SAmple and aggregate (GraphSAGE) model to identify and classify urban commercial spaces. As a graph neural network, an adjacency matrix can be input to address the limitations of convolutional neural networks in processing spatial topological structures by utilizing graph convolution operations to extract the spatial context of urban commercial space scenarios. This approach aggregates information from neighboring nodes within the graph, thereby enhancing classification accuracy [26,32], allowing for the consideration of population data from different time periods. It offers unique advantages, particularly in modeling complex urban network structures.

This study is conducted at the “natural neighborhood” scale using POI data, population density data, and vector-based geospatial data. A “natural neighborhood” refers to urban areas divided by boundaries such as roads, railways, and rivers, which impede pedestrian movement, allowing for finer-grained street-level units to be used for the multidimensional analysis of people, land, and infrastructure. Compared to the traditional grid-based approach, studying the spatial patterns of commercial services at the natural neighborhood scale provides a more accurate reflection of population distribution characteristics, aligning better with human cognition and real-world conditions, thus offering greater practical value [33,34]. By employing geospatial analysis methods and the GraphSAGE model, this study investigates the spatial structure of commercial service facilities within the urban area of Beijing’s Sixth Ring Road and their relationship with population density in different time periods. Additionally, graph node embeddings are generated for classification, enabling a refined description and categorization of the spatial patterns of commercial service facilities at the natural neighborhood scale in Beijing.

2. Materials and Methods

2.1. Materials

2.1.1. Study Area

Beijing is the core city of the Beijing–Tianjin–Hebei urban agglomeration, characterized by diverse commercial functions and a rich composition of business types, with a long history of commercial development [35]. As the capital of the country, it serves as China’s political, economic, and cultural center, covering an administrative area of 16,410.54 square kilometers and comprising 16 districts. The city encompasses various types of commercial services, making it an ideal location for studying urban commercial service patterns. According to data from the Beijing Municipal Bureau of Statistics and the National Bureau of Statistics of China, as of 2014, there are approximately 16.338 million people residing within the Sixth Ring Road, accounting for 75.9% of the city’s total population, highlighting its representativeness for research. Therefore, we have chosen the urban monitoring area within the Sixth Ring Road from the 2023 Urban Land Space Monitoring Project as our study area (Figure 1).

2.1.2. Data

The POI data are provided by the National Platform for Common GeoSpatial Information Services, known as “Map World”. These POIs come with a type-identification code provided by the “Map World” platform for classification purposes. Based on the current planning standards, including the “Planning Standards for Urban Public Service Facilities” (GB50442) and the “Classification of Retail Formats” (GB/T18106-2021), commercial service facility POIs are selected and reclassified (

Table 1. Classification of commercial service facility POIs.

Commercial Format	POI Subclasses	Number	Proportion
Catering services	Chinese restaurants, foreign restaurants, local flavors and snacks, fast food restaurants, leisure catering, bars, cafés, beverage shops, tea houses	42,276	21.52%
Accommodation services	Commercial accommodations, hotels, motels, serviced apartments	7542	3.84%
Shopping services	Convenience stores, supermarkets, shopping malls, home appliance and electronics stores, home building materials markets, Mother and baby products retails, alcohol and tobacco retails, clothing retails, stationery and sports goods retails, pharmaceutical products retails, automobile sales and related products retails, other retails	71,598	36.46%
Financial and insurance services	Securities firms, automated teller machines (ATM), financial services, banks, pawnshops	10,972	5.59%
Sports and leisure services	Golf courses, ski resorts, bowling alleys, tennis courts, badminton courts, sports halls, fishing parks, skating rinks, swimming pools, karaoke television (KTV), fitness centers, chess and card rooms, internet cafés, lottery sales, agritainment, theaters, cinema, amusement parks, resorts, zoos, botanical gardens, aquariums	10,786	5.49%
Business services	Commercial facilities, conference centers, training centers, edifices, business centers, rental services, law firms, travel agencies	12,996	6.62%
Resident services	Home services, domestic services, appliance repair, pet clinics, water delivery stations, laundries, telecommunication business offices, beauty salons, bath and massage facilities, logistics and express delivery facilities, photography and printing facilities, post offices, wedding services, real estate agencies	40,223	20.48%

).

Vector geospatial data are sourced from the 2022 National Land Change Survey conducted by the National Geomatics Center of China, used to calculate the proportion of commercial service facility land and urban residential land in each natural neighborhood. Utilizing ArcMap 10.7 software for spatial statistics, the proportions of commercial service facility land and urban residential land in each natural neighborhood are classified into eight categories using the natural breaks method (Figure 2).

Population density data are sourced from the national key research and development project “Geographic Big Data Mining and Spatiotemporal Pattern Discovery” (2017YFB0503600). This dataset was derived from processed operator data collected over ten days in two months of 2018. After spatiotemporal residence reconstruction, real-time population density data were obtained, which calculates the number of mobile users present in each 500 m area on an hourly basis. The temporal resolution is hourly, while the spatial resolution is 500 m.

Based on individual user stay behaviors, the population density data for different time periods can be categorized [36]: the population density during the nighttime period (1:00–6:00) reflects residential distribution; the population density during meal periods (12:00–14:00, 17:00–19:00) indicates non-residential activity hotspots, primarily around dining and related commercial service facilities; and the population density during other time periods reflects general individual activity locations. The population distribution for each neighborhood was obtained through interpolation to calculate average values (Figure 3). These data can partially reflect the relative population density across different neighborhoods.

2.2. Methods

This study proposes an analysis method based on multisource spatiotemporal data and the GraphSAGE model, aimed at revealing the spatiotemporal relationships between commercial service facilities, population density, and land use. The workflow of this method is shown in Figure 4. First, the geographic vector data were preprocessed, involving clipping and overlay analysis. Simultaneously, the POI data were deduplicated and categorized, and the population density data were processed by time periods. Based on this, GIS (Geographic Information System) analysis was employed to systematically explore the relationship between the spatial distribution of commercial service facilities and population density. Finally, an undirected graph G (V, E) of natural neighborhoods was constructed, and graph G was used as an adjacency matrix to represent the geographic spatial relationships between natural neighborhoods. These features, combined with the POI kernel density characteristics and aggregated neighborhood attributes, were input into the GraphSAGE model to generate graph node embeddings. After clustering, the result provides a refined classification of the commercial service patterns in Beijing’s natural neighborhoods.

2.2.1. Average Nearest Neighbor Distance

The average nearest neighbor distance method assesses the overall spatial clustering level of point features by analyzing the ratio (R) of the observed nearest neighbor distance to the expected distance. The results reflect the global spatial clustering state of point features within the study area. When R < 1, the point features are spatially clustered, with smaller values of R indicating a higher degree of clustering. Conversely, when R > 1, the point features are spatially dispersed. If R = 1, the point features are randomly distributed. The significance of the results can be tested using a z-statistic. The calculation formula is as follows:

$$R={\displaystyle \frac{\overline{d_i}}{{\overline{d}}_e}}$$

(1)

In the formula, $\overline{d_i}$ represents the observed average nearest neighbor distance between point features and ${\overline{d}}_e$ is the expected average distance between features under a random distribution.

The standard deviation of $R$(Z) is given by

$$Z=\left(\overline{d_i}-{\overline{d}}_e\right)\sqrt{{\displaystyle \frac{N^2}{A}}}/0.26136$$

(2)

In the formula, $A$ represents the area of the minimum enclosing rectangle that contains all the features and $N$ is the total number of features.

2.2.2. Kernel Density Estimation

The Kernel Density Estimation (KDE) method is a non-parametric estimation technique used to calculate the density of features (POI) within their surrounding neighborhood. The results produced by KDE are smoothly distributed, providing a better reflection of the distance decay pattern [28]. By using the Kernel Density Estimation (KDE) method, this study examines the spatial variation in the density of commercial service POI data, exploring the spatial characteristics of commercial service facilities within specific regions. A quartic polynomial kernel density function was applied in this study, as shown below:

$${\widehat{\lambda }}_h\left(p\right)={\displaystyle \frac{1}{{\left(radius\right)}^2}}\sum _{i=1}^n{\displaystyle \frac{3}{\pi }}{\left[1-{\left({\displaystyle \frac{P-P_i}{radius}}\right)}^2\right]}^2$$

(3)

In the equation, $p$ represents the location to be estimated, ${\widehat{\lambda }}_h\left(p\right)$ is the kernel density estimate at location $p$, and $radius$ is the radius (bandwidth) centered at $p$. The index $i$ = 1,2,…,n denotes the POI points within the bandwidth range and ($P-P_i$) represents the distance between the estimated point $p$ and the $i$-th POI, provided that this distance is less than $radius$.

2.2.3. Spatial Autocorrelation

Moran’s I index serves as an analytical tool for detecting the spatial correlation of various factors, assessing the overall distribution of spatial variables, and examining their clustering characteristics [37,38]. It typically refers to the global Moran’s I (global spatial autocorrelation) [39] and local Moran’s I (local spatial autocorrelation) [40]. Global spatial autocorrelation tests for the presence of any clustering, dispersion, or randomness in space by describing the spatial characteristics of a phenomenon or attribute value across the entire study area [9]. In contrast, local spatial autocorrelation provides a measure for analyzing spatial data in different regions or units within the study area, reflecting the degree and significance of spatial differences between each region or unit and its surrounding areas.

The spatial distribution of urban commercial service facilities is closely related to human activities. Considering this distribution characteristic, this study employs a global spatial autocorrelation model to describe the distribution of human activity as well as the proportions of commercial and residential land use. The calculation formula is as follows

$$Moran’sI={\displaystyle \frac{n{\displaystyle \sum _{i=1}^n}{\displaystyle \sum _{j=1}^n}{W}_{ij}\left(y_i-\overline{y}\right)\left(y_j-\overline{y}\right)}{{\displaystyle \sum _{i=1}^n}{\displaystyle \sum _{j=1}^n}{W}_{ij}{\displaystyle \sum _{i=1}^n}{\left(y_i-\overline{y}\right)}^2}},\left(i\ne j\right)$$

(4)

Once the results of the global autocorrelation analysis indicate a significant correlation, a local autocorrelation analysis can be conducted to examine the correlation of a particular phenomenon within regional units in local space, specifically the similarity of that phenomenon in a unit to its neighboring units. The calculation formula is as follows

$$LocalMoran’sI={\displaystyle \frac{n\left(y_i-\overline{y}\right){\displaystyle \sum _{j=1}^n}{W}_{ij}\left(y_j-\overline{y}\right)}{{\displaystyle \sum _{i=1}^n}{\left(y_i-\overline{y}\right)}^2}},\left(i\ne j\right)$$

(5)

In Equations (4) and (5), $n$ represents the number of spatial units, $W_{ij}$ is the weight matrix measuring the adjacency relationships between the spatial units, $y_i$ and $y_j$ are the attribute values of spatial units $i$ and $j$, and $\overline{y}$ is the average attribute value of the spatial units.

To explore the spatial correlation between commercial service facilities and human activities, a local bivariate spatial autocorrelation model is introduced. The calculation formula is as follows

$$I={\displaystyle \frac{n{\displaystyle \sum _{i=1}^n}{\displaystyle \sum _{j=1}^n}{W}_{ij}\left(y_i^p-{\overline{y}}_p\right)\left(y_j^q-{\overline{y}}_q\right)}{{\displaystyle \sum _{i=1}^n}{\displaystyle \sum _{j=1}^n}{W}_{ij}{\displaystyle \sum _{i=1}^n}\left(y_i^p-{\overline{y}}_p\right)\left(y_j^q-{\overline{y}}_q\right)}},\left(i\ne j\right)$$

(6)

In the equation, $I$ represents the local bivariate Moran’s I index, $n$ is the number of spatial units, $W_{ij}$ is the weight matrix measuring the adjacency relationships between the spatial units, $y_i^p$ is the value of attribute $p$ for spatial unit $i$, $y_j^q$ is the value of attribute $q$ for the neighboring spatial unit $j$ of unit $i$, and ${\overline{y}}_p$ and ${\overline{y}}_q$ are the average values of attributes $p$ and $q$, respectively.

2.2.4. Construction of Graph Structure and Node Features

To further capture the spatial adjacency relationships between analysis units, we introduce the concept of spatial neighbors from spatial information science [41]. If two objects share a boundary, they are considered first-order neighbors, indicating topological adjacency. This study focuses on natural neighborhoods, which are not separated by transportation elements or water systems; thus, each natural neighborhood exists independently in space.

In this study, an undirected graph G (V, E) of natural neighborhoods was constructed by taking the blocks as graph nodes V and lines between the blocks with proximity as edges E. We used a constrained Delaunay triangulation to determine the proximity of two natural neighborhoods. By calculating the average radius of the circumscribed circle of the natural neighborhoods, which is 252.166 m, we establish a complete constrained Delaunay triangulation network by limiting the edge length L_max of the Delaunay triangles to 1000 m. When two natural neighborhoods share a triangle edge, an adjacency relationship (E) is established between the corresponding graph nodes (V) (Figure 5). The constructed undirected graph G was input into the GraphSAGE model as an adjacency matrix to represent the actual geographic spatial relationships.

In selecting features for each node (V), we chose the number of commercial service facilities to represent the commercial development status of the neighborhood itself. The kernel density values of various commercial service facility points of interest (POIs) were employed to characterize the coverage of these facilities. The proportions of two types of land use (land for commercial service facilities and land for urban residential purposes) in each neighborhood were used to describe the urban functional zoning of the block. Furthermore, population density data for nighttime (1:00–6:00) and meal periods (12:00–14:00, 17:00–19:00) were utilized to simulate the associations between commercial activities, residential activities, and population movements.

2.2.5. Classifying Natural Neighborhoods Using the GraphSAGE Model

Graph Neural Networks (GNNs), as one of the methods for knowledge graph embedding representation, can effectively learn features from graph structures and node attributes, demonstrating excellent performance in generating low-dimensional vectors that represent nodes and edges. The GraphSAGE model belongs to the inductive learning category of GNN algorithms, where the feature learning process for each node is not only related to the node’s own attributes but also aggregates features from neighboring nodes. It only considers the information from local neighbors without the need to traverse the entire graph, allowing the model to exhibit strong generalization capabilities and make predictions even when the graph structure changes [32]. In hyperparameter selection, the depth of network search significantly influences the aggregation and representation of node features within a graph [42]. When the search depth is set to two layers, the precision of results improves by 10% to 15% compared to a single layer. However, increasing the depth beyond two layers yields only marginal gains in precision (0% to 5%) while significantly escalating computational time, which increases by a factor of ten to one hundred [32]. As a result, a two-layer search depth is commonly selected as the default hyperparameter in many studies [43,44].

The basic operational steps of the GraphSAGE model are as follows:

• Determine the number of layers for the network search depth. For instance, with $k$ = 2, each node can learn its own embedding based on the features of its second-order neighbor nodes.
• Perform aggregation operations in each layer of the loop; select an appropriate model operator (aggregator) to output the overall features of each node’s neighbors in the graph. During this process, the aggregated neighbor features are concatenated with the previous layer’s state information of the selected node, and then the computed node features are normalized.
• Obtain the embeddings of each node in the graph after $n$ iterations.

This study employs an unsupervised GraphSAGE model, with the loss function based on the negative sampling method:

$$J_G\left(Z_u\right)=-\mathit{log}\left(\sigma \left(Z_u^T{z}_{\nu }\right)\right)-Q\cdot E_{{\nu }_n~P_{n\left(v\right)}}\log \left(-\sigma \left(Z_u^T{z}_{\nu }\right)\right)$$

(7)

In the equation, $Z_u$ represents the embedding generated for node $u$ through the GraphSAGE model; node $v$ is a “neighbor” that node $u$ reaches via random walks; $\sigma$ is the sigmoid function; $P_n$ is the probability distribution for negative sampling, similar to the negative sampling in Word2Vec [45]; and $Q$ is the number of negative samples.

Using the GraphSAGE model to identify and classify commercial service patterns offers significant advantages. Urban POI data collection often suffers from data deficiencies in certain areas due to technical limitations, resource shortages, or delays in data updates. GraphSAGE predicts the features of target nodes by sampling and aggregating the characteristics of neighboring nodes [27], thus reducing reliance on a complete feature set and effectively overcoming this issue. When evaluating commercial service patterns, it is crucial to consider the influence of surrounding neighborhoods, as commercial service patterns are shaped by the cumulative effects of the surrounding environment. GraphSAGE can capture the genuine interaction relationships between adjacent geographic features, allowing for more accurate differentiation of target nodes and interpretation of their attributes, thereby enhancing classification accuracy.

2.2.6. Commercial Service Patterns Clustering

Based on the vector representations generated by GraphSAGE, further clustering analysis is conducted. Clustering analysis is an unsupervised classification process aimed at grouping similar patches into the same cluster, thereby identifying areas with similar urban commercial service patterns. In this study, the k-means algorithm is employed for the clustering analysis.

External and internal testing methods are used to evaluate the clustering results. The external testing method assesses the validity of the clustering approach by comparing the clustering results with the true classification outcomes of the dataset. In contrast, the internal testing method relies solely on the dataset itself to evaluate the accuracy of the clustering results. When the dataset lacks true labels, the internal testing method serves as a criterion for setting the model training parameters. We employ two internal testing methods to determine the optimal number of clusters: the “elbow method,” which calculates the sum of squared errors (SSE) [46], and the calculation of the silhouette coefficient (SIL) [47].

$$SSE=\sum _{i=1}^m\sum _{x_i\in {\Omega }_i}{\left|x_i-{\overline{x}}_i\right|}^2$$

(8)

In the equation, $m$ represents the number of clusters, ${\Omega }_i$ denotes all units that are assigned to the $i$-th cluster, $x_i$ is the representation vector of the unit, and ${\overline{x}}_i$ is the average value of all units in the $i$-th cluster.

$$\mathrm{S}IL={\displaystyle \frac{1}{n}}\sum _{j=1}^n{\displaystyle \frac{a-b}{\max (a,b)}}$$

(9)

In the equation, $n$ is the number of data points, $a$ represents the average distance between the $j$-th data point and the other data points in its assigned cluster, reflecting the compactness (cohesion) of the cluster, while $b$ denotes the average distance between the $j$-th data point and the data points in the nearest cluster.

3. Results and Analysis

3.1. Spatial Distribution Characteristics

The average nearest neighbor ratios for all seven categories of commercial service facilities are less than 1 (

Table 2. Average nearest neighbor analysis of commercial service facilities in Beijing.

Commercial Format	Average Nearest Neighbor Distance (m)	Nearest Neighbor Ratio	Z-Score	p-Value	Clustering Degree Ranking
Financial and insurance services	45.61	0.185	−163.28	0.00	1
Catering service	25.06	0.198	−315.41	0.00	2
Shopping services	20.05	0.206	−406.66	0.00	3
Accommodation services	93.82	0.317	−113.42	0.00	4
Resident services	42.27	0.323	−259.69	0.00	5
Business services	73.18	0.325	−147.18	0.00	6
Sports and leisure services	105.57	0.423	−114.46	0.00	7

). The degree of clustering, from highest to lowest, is as follows: financial services > dining services > shopping services > accommodation services > residential services > business services > sports and leisure services. The nearest neighbor ratios for financial services, dining services, and shopping services are lower than the average ratio for all commercial types (0.282), indicating that commercial activities with high consumption frequency, such as dining and shopping services, exhibit higher spatial clustering. The particularly low nearest neighbor ratio for financial services may be due to the specialized nature of the financial industry, where banking and insurance service outlets tend to be concentrated in specific locations. In contrast, services like sports and leisure facilities, some of which are guided by the government, emphasize spatial equity in their distribution, resulting in the lowest degree of clustering among the seven categories of commercial services.

Using the Kernel Density Estimation (KDE) tool in ArcMap, a kernel density analysis was conducted on the points of interest (POI) data for commercial service facilities within the urban monitoring area inside Beijing’s Sixth Ring Road. Considering the service radius of commercial service facilities, the search radius for the kernel density analysis was set to 1000 m. The results of the analysis were classified into six categories using the natural breaks (Jenks) classification method (Figure 6).

The kernel density values for overall commercial service facilities (Figure 6a) range from 0 to 2283.875, displaying distinct hierarchical spatial distribution characteristics. Within the Second Ring Road (Dongcheng District and Xicheng District), the density of commercial service facilities is significant, exhibiting strong central clustering features. Between the Second Ring Road and the Fourth Ring Road (including Chaoyang District, Haidian District, Fengtai District, and Shijingshan District), this area shows a contiguous clustering pattern, particularly on the eastern side, where commercial conditions are markedly better than in other parts. Between the Fourth Ring Road and the Fifth Ring Road, a differentiation occurs, with the western side displaying a patchy dispersed pattern, while the eastern side continues to maintain contiguous clustering with a relatively higher level of commercial services. Finally, in areas beyond the Fifth Ring Road, commercial service facilities exhibit a decentralized multi-center pattern with lower density. This spatial distribution reflects a trend from the core areas to the peripheral regions, where commercial service facilities gradually shift from high-density clustering to a multi-centered dispersion.

The spatial distribution of various commercial service facilities exhibits distinct differentiation. Specifically, the spatial distribution of dining service facilities (Figure 6b) closely resembles that of overall commercial service facilities, while accommodation service facilities (Figure 6c) are predominantly concentrated within the Fourth Ring Road in Beijing. The eastern side within the Second Ring Road demonstrates primary and secondary high-density clustering, particularly in the Wangfujing and Qianmen Dashilan commercial areas, where other commercial service facilities also exhibit high density due to their proximity to Beijing’s famous landmarks and convenient transportation. Consequently, the spatial distribution of accommodation facilities is influenced by factors such as transportation infrastructure and tourist attractions.

Shopping service facilities (Figure 6d) display a significant strong centrality, with two major commercial areas, Xidan and Wangfujing, located within the Second Ring Road, surrounded by contiguous clusters of primary, secondary, and tertiary densities. Additionally, the Gongzhufen commercial area is situated within the central city (inside the Fourth Ring Road), but it does not exhibit contiguous clustering around it, instead reflecting a strong single-point centrality. The spatial distribution of financial and insurance service facilities (Figure 6e) is closely tied to the planning of Beijing’s central business district, particularly exhibiting primary and secondary high-density clustering in the Guomao CBD area, as well as in Xidan Financial Street and Zhongguancun in Haidian District.

Sports and leisure service facilities (Figure 6f) are primarily concentrated in Sanlitun, located between the Second and Third Ring Roads, exhibiting primary and second-order high-density clustering, while other regions display multi-centered and dispersed patterns. The spatial distribution of business service facilities (Figure 6g) is similar to that of financial and insurance services, closely associated with the planning of Beijing’s central business district. Finally, the spatial distribution of residential service facilities (Figure 6h) shows contiguous clustering in areas west of the Second, Third, and Fourth Ring Roads, whereas other regions exhibit multi-centered and dispersed patterns.

Overall, the spatial distribution of various commercial service facilities in Beijing exhibits a distinct geographic pattern, characterized by clustering features that create a multi-tiered layout of central aggregation and peripheral dispersion. The distribution of these facilities is significantly influenced by factors such as transportation, tourist attractions, and the planning of central business districts. Additionally, commercial service facilities on the west side of the central axis perform better overall than those on the east side, reflecting the uneven development characteristics of the region.

3.2. Spatial Autocorrelation Analysis

The global Moran’s I index provides an overall assessment of spatial autocorrelation characteristics.

Table 3. Global Moran’s I results.

Variable	Global Moran’s I	Standard Deviation	Z-Score	p-Value
Commercial service land	0.2515	0.004	63.4352	p < 0.001
Urban residential land	0.4040	0.004	100.4621	p < 0.001
Population density (1:00–6:00)	0.3265	0.004	82.1156	p < 0.001
Population density (12:00–14:00, 17:00–19:00)	0.3244	0.0039	83.3059	p < 0.001

displays the global Moran’s I indices for the proportion of land allocated to commercial services, the proportion of urban residential land, population density (1:00–6:00), and population density (12:00–14:00 and 17:00–19:00). The results indicate that all Moran’s I values are greater than 0, suggesting the presence of positive global spatial autocorrelation characteristics. Furthermore, at a 99.9% confidence level, all indicators exhibit p-values less than 0.001, confirming their significance in the statistical tests (Table 3).

The global Moran’s I index only examines overall spatial autocorrelation. To further detect specific types of spatial autocorrelation between individual natural neighborhoods, the local Moran’s I test is employed to reveal spatial clustering characteristics and confirm clustering effects.

The H-H clusters of commercial service facility land within the Second Ring Road (Figure 7a) are primarily concentrated on the east, west, and south sides of the Forbidden City, corresponding to the Wangfujing, Xidan, and Qianmen Dashilan commercial areas. In contrast, the L-L clusters are concentrated around the Forbidden City and its surrounding cultural heritage sites and parks. Between the Second and Fourth Ring Roads, H-H clusters are mainly located on the east, southwest, and northwest sides, corresponding to Guomao, Lize Business District, and Zhongguancun, respectively. Between the Fourth and Fifth Ring Roads, H-H clusters are concentrated in the southwest and northeast, while areas beyond the Fifth Ring Road are more evenly distributed. Some H-H clusters are present in the Xibeiwang commercial area and the northeast side of the Fifth Ring Road, with widespread contiguous L-L clusters.

The distribution of urban residential land (Figure 7b) is relatively uniform. Within the Second Ring Road, H-H clusters dominate, primarily concentrated in hutongs and historical cultural districts, while L-H outliers are located near the Forbidden City and surrounding parks. Between the Second and Fourth Ring Roads, H-H clusters are concentrated in the southern and western areas, with fewer L-H and H-L outliers. Beyond the Fourth Ring Road, H-H clusters are scattered, while L-L clusters exhibit widespread contiguous distribution.

The local Moran’s I for population density during the nighttime period (1:00–6:00) (Figure 8a) shows that H-H clusters are densely concentrated between the Second and Fourth Ring Roads, with additional H-H clusters present in the northern and eastern areas outside the Fourth Ring Road. L-L clusters are primarily located around the cultural heritage sites and parks near the Forbidden City within the Fourth Ring Road, as well as in the southwest Lize Business District. Outside the Fifth Ring Road, L-L clusters dominate, with scattered H-L outliers within the Second Ring Road.

For the meal periods (12:00–14:00, 17:00–19:00) (Figure 8b), the population density shows that H-H clusters are concentrated between the Second and Fourth Ring Roads, while L-L clusters are mainly found around the Forbidden City and the Lize Business District within the Third Ring Road. H-L outliers are more scattered within the Second Ring Road, and there are many L-L clusters outside the Fifth Ring Road. Comparing the two time periods, the number of H-H clusters outside the Fifth Ring Road decreases during the dinner periods compared to the nighttime, indicating a trend of population clustering toward the city center, suggesting that the city center is more attractive during the daytime.

The bivariate global spatial autocorrelation shows that the Moran’s I indices between the kernel density values of all seven types of commercial service facilities and population density are greater than 0, indicating positive global spatial autocorrelation. At a 99.9% confidence level, all p-values are less than 0.001, passing the significance test (

Table 4. Bivariate global Moran’s I results.

Commercial Format	Catering Service	Accommodation Services	Shopping Services	Financial and Insurance Services	Sports and leisure Services	Business Services	Resident Services
Population density (1:00–6:00)	0.1625 (53.98) ***	0.0908 (30.81) ***	0.1167 (39.39) ***	0.2011 (65.17) ***	0.3087 (96.17) ***	0.1766 (57.98) ***	0.2901 (91.85) ***
Population density (12:00–14:00, 17:00–19:00)	0.2969 (94.89) ***	0.1790 (59.77) ***	0.2082 (68.71) ***	0.3776 (116.58) ***	0.3837 (117.17) ***	0.3799 (117.99) ***	0.3310 (104.39) ***

Note: the values within parentheses are z-scores; *** represents p < 0.001.

). The Moran’s I indices between population density and various types of commercial service facilities during the meal periods (12:00–14:00, 17:00–19:00) are higher than those during the nighttime period (1:00–6:00). This suggests that, during meal periods, the spatial distribution of population density is more strongly correlated with the distribution of commercial service facilities.

Furthermore, we conducted a bivariate local spatial autocorrelation analysis on population density and the kernel density of seven types of commercial service facilities, revealing significant spatial heterogeneity. The analysis results for the nighttime period (1:00–6:00) (Figure 9) and the meal periods (12:00–14:00, 17:00–19:00) (Figure 10) indicate that H-H clusters are primarily concentrated within the Fifth Ring Road, while L-L clusters are distributed in areas outside the Fourth Ring Road. H-L outliers are predominantly found in regions beyond the Third Ring Road, whereas L-H outliers exhibit the opposite pattern, with a higher frequency of occurrence as proximity to the city center increases.

The H-H clusters of dining service facilities (a) are primarily concentrated within the Fifth Ring Road, with contiguous aggregation observed on the eastern side, while the remaining areas exhibit a multi-centered dispersed distribution, with numerous L-H outliers present within the Second Ring Road. The H-H clusters of accommodation service facilities (b) are mainly located within the Third Ring Road, expanding outward from the center around the Forbidden City. For shopping service facilities (c), H-H clusters are predominantly found within the Fourth Ring Road, with contiguous aggregation observed between the Fourth and Third Ring Roads on the eastern side, while other areas show a multi-centered dispersed pattern. The H-H clusters of financial and insurance service facilities (d) are mainly concentrated within the Fourth Ring Road, particularly in contiguous clusters on the northwest and eastern sides. Similarly, the H-H clusters of sports and leisure service facilities (e) are primarily distributed within the Fourth Ring Road, with a notable concentration of H-H clusters in the northeastern region.

The L-H and H-L outliers for business service facilities (f) significantly decrease during meal periods compared to the nighttime period, likely due to the fact that these facilities primarily consist of office spaces, where population density increases during mealtimes but decreases during nighttime. Furthermore, some neighborhoods transition from H-H clusters to L-H outliers from meal periods to nighttime, indicating that business service facilities are more concentrated in areas beyond the Fourth Ring Road, while changes within the Fourth Ring Road primarily occur near Xidan Financial Street and the Guomao area. H-H clusters of residential service facilities (g) are predominantly located within the Fifth Ring Road, exhibiting contiguous aggregation on the eastern side, with other areas displaying a multi-centered dispersed distribution.

Combined with the changes observed in the bivariate global Moran’s I indices, the Moran’s I index for business service facilities increased from 0.1766 during the nighttime period to 0.3799 during the meal periods, indicating a stronger spatial tendency to concentrate in areas of high population density during the meal periods, particularly along and within the Fourth Ring Road, which are hotspots for population movement and aggregation. In contrast, the Moran’s I index for residential service facilities only showed a slight increase from 0.2901 to 0.3310, representing the smallest growth, suggesting that their spatial distribution is less influenced by population activities and exhibits greater stability. This may be attributed to the close integration of residential service facilities with daily life; even beyond the Fifth Ring Road, residential service facilities remain widely distributed in neighborhoods with a high concentration of residential land, catering to the everyday needs of residents.

Considering the density characteristics and changes in the Moran’s I indices discussed above, we can infer that the high concentration of business service facilities is closely related to the population density in specific areas during meal periods, reflecting a temporal and spatial alignment between economic activities and population movement. In contrast, residential service facilities, due to the ubiquitous nature of their services, exhibit greater balance and stability in their spatial distribution, spanning different layers and functional zones to ensure the convenience of residents’ daily lives.

3.3. Description and Classification of Commercial Service Patterns Based on GraphSAGE

3.3.1. Graph Node Embedding Generating and Clustering

The previously extracted node semantic and spatial features were fed into the GraphSAGE model to generate node feature embeddings. Subsequently, the K-means algorithm was used for clustering analysis of the node embeddings, and the silhouette coefficient and elbow method were applied to evaluate the separation of the clustering results (Figure 11), ultimately selecting six clusters as the optimal classification number. The TSNE (T-Distributed Stochastic Neighbor Embedding) visualization of the node embeddings shows a clear clustering pattern among the categories (Figure 12).

A confusion matrix was implemented to evaluate the accuracy of nature neighborhoods classification. Twenty blocks in each of the six types of clusters were selected randomly. Then, the clustering results were manually verified using satellite imagery and annotated maps from the Baidu Maps app. The evaluation results are listed in

Table 5. Confusion matrix results of classification.

	Cluster 1	Cluster 2	Cluster 3	Cluster 4	Cluster 5	Cluster 6	Total
Cluster 1	15	1	0	0	3	1	20
Cluster 2	0	16	0	4	0	0	20
Cluster 3	0	0	19	1	0	0	20
Cluster 4	0	1	2	17	0	0	20
Cluster 5	0	0	0	0	18	2	20
Cluster 6	6	0	0	0	0	14	20
Total	21	18	21	22	21	17	120

. The overall accuracy (OA) of clustering was 82.5%, and the kappa coefficient was 0.79.

To evaluate the performance of classification, the used model was compared with commonly used models, including K-means [4], Latent Dirichlet Allocation (LDA) [48], and Agglomerative Nesting (AGNES) [49,50]. In total, 120 manually identified units were selected for each comparison (

Table 6. Overall accuracies and kappa coefficients of the classification models.

Index	Model	OA	Kappa
1	LDA	44.17%	0.33
2	K-means	58.33%	0.50
3	AGNES	49.17%	0.39
4	GraphSAGE	82.50%	0.79

). The experimental results indicate that integrating spatial relationships into attribute feature extraction more effectively captures the characteristics of blocks. This approach shows a significant improvement compared to the direct use of the K-means clustering algorithm, which enhances the classification performance of commercial service patterns.

3.3.2. Classification Results of Commercial Service Patterns

Mapping the clustering results of node embeddings onto the natural neighborhood graph (Figure 13) reveals, to some extent, the spatial patterns of commercial service facilities in urban Beijing. The spatial distribution of different categories of natural neighborhoods reflects the clustering and diffusion characteristics of commercial service facilities, thereby illustrating the geographic differentiation of various commercial functional areas. This mapping not only aids in intuitively understanding the distribution patterns of commercial service facilities within the city, but also provides a basis for further analysis of the spatial relationships within urban functional areas.

Figure 14 illustrates the clustering results of node embeddings in relation to various features of natural neighborhoods, including the proportion of land allocated to commercial service facilities, the proportion of urban residential land, population density (1:00–6:00), and population density during meal periods (12:00–14:00, 17:00–19:00). Additionally, it encompasses the quantity and density of different types of POIs. The combination of these features provides important insights into understanding the characteristics of natural neighborhoods concerning the distribution of commercial service facilities, residents’ living environments, and population activity patterns. This further reveals the spatial structure and functional zoning of commercial service facilities in urban Beijing.

Cluster 1: Historical and cultural districts in the center of Beijing. This cluster is primarily located within the Second Ring Road, adjacent to tourist attractions such as the Forbidden City, Temple of Heaven, Beihai Park, and Jingshan Park. In this cluster, population density is relatively low, and there is an extremely high proportion of urban residential land, which results in limitations on commercial development. However, due to its geographic location, the commercial service coverage remains quite favorable.

Cluster 2: Natural neighborhoods located on the urban fringe. This cluster is primarily situated in areas outside the Fourth Ring Road of Beijing, with many natural neighborhoods between the Fourth and Fifth Ring Roads consisting mainly of parks and green spaces. These natural neighborhoods exhibit the lowest population density, proportion of land allocated to commercial service facilities, and proportion of urban residential land among all natural neighborhoods, indicating a lack of objective conditions for commercial development. Additionally, due to their geographic location on the outskirts of the city, the coverage of commercial services in these areas is relatively poor.

Cluster 3: The spatial distribution of this cluster resembles a transitional zone between Cluster 4 and Cluster 5. In terms of characteristics, these neighborhoods have a relatively high proportion of urban residential land, while also allocating an appropriate portion of land for commercial service facilities, indicating potential for commercial development. The population density in this cluster is higher than the average compared to other clusters, and the coverage of residential service facilities and sports and leisure services meets the basic needs of the residents.

Cluster 4: This cluster is characterized as a transitional zone between the suburbs and urban areas, primarily situated in regions beyond the Fourth Ring Road in Beijing, extending inward to the Third Ring Road. It includes prominent landmarks and parks such as the Forbidden City, Temple of Heaven, and Yuyuantan Park. The proportion of commercial service facility land and urban residential land in these neighborhoods is relatively low, leading to limited commercial development. Furthermore, despite the considerable population residing in these areas, the coverage of commercial services remains insufficient.

Cluster 5: This cluster is characterized by its highest residential density, greatest number of residential service facilities, and excellent coverage of commercial services. It is primarily located between the Second and Fourth Ring Roads, as well as on the northern and eastern sides between the Fourth and Fifth Ring Roads, while other areas display a more dispersed, point-like layout. The proportion of urban residential land ranks second among all clusters, with the highest population density observed during the nighttime period (1:00–6:00). Notably, high population densities are also observed during mealtimes (12:00–14:00, 17:00–19:00), and commercial service facilities are well-developed.

Cluster 6: This cluster is characterized as a commercial center by its highest proportion of land allocated to commercial service facilities, primarily located within the Fifth Ring Road. Representative districts in this cluster include Xidan Financial Street, Wangfujing, Qianmen Dashilan, Guomao CBD, Zhongguancun, and Sanlitun. The population density during meal periods (12:00–14:00, 17:00–19:00) is significantly higher than during the nighttime period (1:00–6:00), and the proportion of land allocated to commercial service facilities exceeds that of urban residential land.

In general, the relationships among the various types of natural neighborhoods can be summarized as follows. Within the Second Ring Road, the area primarily consists of historical and cultural neighborhoods. Although commercial development is restricted to protect historical sites, the surrounding major commercial districts ensure extensive coverage of commercial services. During meal periods (12:00–14:00, 17:00–19:00), while the population density in this area is not high, commercial activities remain vibrant due to the attractiveness of the surrounding commercial districts. The region between the Second and Fourth Ring Roads exhibits a high degree of integration between residential and commercial areas, with population density leading among all neighborhoods, maintaining a relatively balanced distribution between meal periods and nighttime period (1:00–6:00), thus becoming a core residential and consumption area in the city. The transitional characteristics of the area between the Fourth and Fifth Ring Roads are particularly pronounced, with some regions designated as ecological green spaces and relatively limited commercial coverage. In contrast, other areas serve as mixed-use developments that generally meet residents’ needs, albeit with a more scattered distribution of commercial districts, resulting in localized consumption zones. Beyond the Fifth Ring Road, the area primarily consists of newly expanded urban regions, characterized by overall lower population density and commercial development levels, with only a few regions displaying better integration of residential and commercial features, where commercial activities are relatively weak during meal periods.

At the same time, the spatial relationship between commercial service facilities, population distribution, and urban land planning exhibits significant differences in Beijing’s northern and southern regions, as divided by Chang’an Avenue, and in the eastern and western regions, as divided by the Forbidden City. In the northern region, residential density is significantly higher than in the southern region, and the density of commercial service facilities is also relatively high, with their number clearly exceeding that of the southern region. In contrast, the eastern region has a significantly greater number of commercial center neighborhoods compared to the western region, and the distribution of commercial service facility density reflects this trend, indicating a more concentrated characteristic of commercial facilities in the eastern region. This spatial disparity not only reveals the asymmetry in urban function allocation and service provision across different areas of Beijing but also reflects the diversity and complexity inherent in the urban development process.

4. Discussion

The spatial imbalance of urban functions is a common phenomenon in most cities globally. This imbalance typically manifests in the concentration of cultural and educational resources [15], medical services [51], and infrastructure in the central urban areas [52,53], while peripheral areas are often designated for residential, industrial, and other functions, lacking the necessary commercial and service facilities [54]. According to urban economics theories, the Alonso Model [55] explains the differentiation of land use functions between city centers and peripheral areas, a spatial division that further leads to significant disparities in development potential across different regions. Against this backdrop, cities have gradually adopted a “ core-periphery” model [56]. The core urban area becomes a hub for the concentration of resources, capital, information, and power, while the peripheral areas remain in a passive position in terms of resource allocation, serving the economic activities of the core urban area. Due to insufficient infrastructure investment in the peripheral regions, they become economically dependent on the central areas and face challenges in cultivating independent growth points within their own boundaries.

From the perspective of urban geography, this paper provides a detailed analysis of the spatiotemporal relationships between commercial service facilities, population distribution, and urban land planning in Beijing, revealing significant spatial heterogeneity. The experimental results indicate substantial differences in the distribution of commercial service facilities and population density across different natural neighborhoods in Beijing. The urban core area, benefiting from its resource agglomeration effect, not only attracts a large number of high-income residents but also establishes a well-developed network of commercial and service facilities. In contrast, the peripheral areas, particularly those beyond the Fifth Ring Road, experience lagging commercial development and a lack of necessary commercial services and infrastructure, leading to significant disparities in economic and life opportunities for residents compared to those in the core areas.

Further analysis reveals that the high-density commercial service coverage in the urban core areas aligns with the high population density during meal periods, reflecting strong commercial attractiveness. In the peripheral areas, although some neighborhoods exhibit high population density during nighttime hours, indicating high residential density, the lack of adequate commercial facilities forces residents to travel to the urban core areas for work and consumption, further exacerbating the imbalance in resource flow. This phenomenon aligns closely with the classic “center-periphery” theory, demonstrating how the core areas gain advantages through resource concentration and economic absorption, while the periphery serves as a satellite for the urban core’s economic activities, with its socioeconomic development being significantly constrained.

The excessive concentration of resources may lead to the continuous accumulation of socioeconomic disadvantages in urban peripheral areas, thereby hindering balanced urban development [57]. Taking Beijing as an example, the lagging commercial development and insufficient infrastructure in peripheral areas impede the formation of effective growth points for economic activities. This not only exacerbates socioeconomic inequalities within the city but also poses challenges to sustainable urban development. Therefore, future urban planning must focus on narrowing the gaps between regions by optimizing functional layouts and ensuring the rational allocation of resources to promote inclusive development throughout the city. In light of this, we can propose several recommendations for promoting urban and commercial service development in Beijing from three aspects: the spatial distribution of commercial service facilities, resident distribution, and urban land planning.

• Promoting balanced development inside and outside the Second Ring Road. Kernel density analysis reveals that commercial service facilities within the Second Ring Road are concentrated in areas such as Wangfujing, Xidan, and Qianmen Dashilan, while other regions are predominantly classified as Cluster 1 neighborhoods. These areas are characterized by a limited number of commercial facilities and primarily residential land with low residential density, which restricts the potential for commercial development. The bivariate local Moran’s I index further supports this, indicating the presence of the most L-H clusters within the Second Ring Road, suggesting good coverage of commercial service facilities. However, as shown in Figure 14, there are still gaps in shopping service facilities. Therefore, the government should prioritize the relocation and reorganization of commercial entities to foster the development of surrounding commercial districts, such as Chongwenmen, Dongzhimen, and Xizhimen, thereby promoting the balanced development of consumer hubs. For historical and cultural neighborhoods, efforts should focus on supplementing convenience stores and other commercial formats to enhance the quality of community living circles.
• Promoting commercial development in areas beyond the fifth ring road. Targeted commercial development strategies should be formulated based on population density and the existing commercial infrastructure for emerging urban areas outside the Fifth Ring Road. The research findings indicate that numerous Cluster 3 blocks are distributed in the areas beyond the Fifth Ring Road, characterized by a high population but relatively underdeveloped commercial infrastructure (Cluster 6 blocks are notably absent). Bivariate local Moran’s I analysis further supports this observation, revealing more H-L clusters in these areas, particularly during meal periods (12:00–14:00 and 17:00–19:00) compared to the nighttime period (1:00–6:00). This suggests a trend of population movement towards the inner rings during the day. Consequently, commercial facilities beyond the Fifth Ring Road might not meet the needs of local residents, prompting them to travel inward for shopping and work. To address this issue, the government could implement policies to incentivize the development of large-scale commercial projects and business facilities in densely populated and well-connected areas beyond the Fifth Ring Road, thereby creating attractive commercial nodes. This strategy would not only alleviate functional pressures on urban areas within the Fourth Ring Road and reduce traffic congestion, but also contribute to the balanced development of the city.
• Promoting balanced development across east–west and north–south axes, matching residents’ needs. Significant spatial disparities in commercial development exist between the north and south, as well as the east and west, with Chang’an Avenue dividing the city into northern and southern regions and the Forbidden City separating the eastern and western areas. Specifically, the northern and eastern regions exhibit more developed commercial services, with clustering results showing a higher concentration of Cluster 5 and Cluster 6 neighborhoods, along with a denser population distribution. However, the presence of numerous Cluster 3 neighborhoods in the west, particularly in areas between the Third and Fifth Ring Roads, indicates insufficient coverage of commercial service facilities. Based on regional functions and residents’ needs, the layout of commercial service facilities in this region should be flexibly adjusted. Around ecological green spaces, the development of eco-tourism and leisure services should be pursued moderately to align with green commercial practices. In areas with mixed residential and commercial development, efforts should focus on enhancing the coverage and quality of commercial facilities to meet the diverse consumption needs of residents.

Through the progressive optimization of commercial space configuration, the gap between the urban core and peripheral areas can be narrowed. These approaches not only alleviate commercial pressure on central city areas, promoting the dispersion and balanced allocation of resources, but also address issues related to resource concentration under the “core-periphery” model. By reducing socio-economic inequalities across regions and achieving a more equitable distribution of resources, this strategy can help alleviate urban traffic congestion, foster sustainable urban development, and enhance overall socio-economic conditions. Therefore, future urban planning should prioritize infrastructure development in peripheral areas to enhance their capacity for independent growth, thereby achieving balanced development and social equity across the city.

5. Conclusions and Prospects

In the context of a rapidly developing commercial economy, optimizing the urban commercial spatial structure to meet the increasingly diverse needs of residents has become a key task for enhancing quality of life and promoting economic development. This study, based on multisource urban data, employs geographic spatial analysis methods such as average nearest neighbor distance and kernel density analysis to reveal the spatial distribution characteristics of commercial service facilities in Beijing. Furthermore, by integrating spatial autocorrelation analysis and graph neural network models, this research provides a precise characterization of the commercial service patterns at the natural neighborhood scale from the perspective of spatiotemporal element associations, highlighting the spatial heterogeneity between commercial service facilities, population distribution, and urban land planning in Beijing. This research offers important references for future urban planning and policy formulation, as well as new directions for optimizing the commercial service spatial structures in other major cities. The GraphSAGE model, as employed in this study, embodies an “inductive” algorithmic framework, well-suited for the inductive representation learning of large-scale graph data. It boasts great generalization capabilities and scalability. Moreover, its performance on smaller-scale graph datasets is exemplary, maintaining robust outcomes even when the quantity of node samples is notably limited [58]. However, this model is not without its limitations. Owing to its characteristic as a black box in unsupervised graph neural network modeling, the behavior of the model lacks interpretability, necessitating semantic interpretation of the output results.

With the continuous development and widespread application of big data technology, the identification and description of urban spatial structures will become more precise and comprehensive through the expansion of data sources and the integration of multidimensional urban information. Building on this foundation, future research should further explore the interaction mechanisms between commercial service patterns and factors such as urban planning, transportation networks, and population mobility. Additionally, it is crucial to strengthen the study of the types of commercial service facilities and their service quality to better meet the increasingly diverse consumption needs of residents. Furthermore, considering the rapid growth of e-commerce in the Internet age, an in-depth examination of the impact of online shopping on the urban commercial spatial structure will be an important direction for future research.