High-Vitality Stability Characteristics and Nonlinear Mechanisms of Urban Virtual Vitality: Evidence from Five Urban Districts in Harbin, China

Abstract

Virtual vitality has become an important complementary dimension for describing urban vitality; however, the identification and formation mechanisms of its stable, high-vitality state during dynamic change remain insufficiently explored. Taking five urban districts of Harbin as the study area, this study uses TikTok short-video data from July to August 2024 (summer) and December 2024 to January 2025 (winter), together with Gaode Map POI data, as the core dataset. Kernel density differences between adjacent weeks are used to measure the dynamic changes in virtual vitality. Bivariate local spatial autocorrelation is applied to identify high-vitality stable zones, and a Random Forest model is employed to examine the nonlinear influence of physical vitality spatial structures. The results show the following: (1) Dynamic change patterns of virtual vitality differ significantly across seasons, and when online attention content points to specific physical spatial structures, a stable high-vitality state is more likely to be maintained. (2) Bivariate local spatial autocorrelation analysis indicates that high-vitality stable zones (HH zones) exhibit significant spatial clustering, with vitality-enhancing zones (LH zones) distributed around them and showing spillover effects, while vitality-declining zones (HL zones) are more scattered. (3) The Random Forest results show that the stable maintenance of high virtual vitality depends more on combinations of spatial structural characteristics with high recognizability, among which distance to activity center (tourism), functional composition dissimilarity (culture), and functional composition dissimilarity (shopping) have the strongest influence. These findings reveal a nonlinear relationship between the stable high-vitality state and the structure of physical vitality space, providing insights for guiding online attention to support physical spatial development.

1. Introduction

Urban vitality is usually understood as the intensity of human activities and the level of social interaction in urban space, and it is an important indicator for measuring the quality of urban operation and spatial attractiveness [1]. Traditional studies have mainly taken built environment factors as the core and explained the formation mechanism of physical vitality from aspects such as land use mix, population density, and accessibility [2,3]. With the spread of the internet and social media, urban vitality has gradually shown a dual structure of “offline experience–online expression”. Researchers have begun to use social media behavioral data to represent urban virtual vitality, in order to capture dimensions that physical vitality cannot fully reflect, such as cultural atmosphere, perceived experience, emotional preference, and city image [4]. Early studies mainly defined virtual vitality as a dimension of virtual accessibility embedded in physical space, which was used to represent the basic ability of urban space to connect with digital space [5]. As research has continued to deepen, scholars have further emphasized its symbolic and representational characteristics, arguing that virtual vitality is not only an online projection of physical activities but is also jointly influenced by narrative expression and emotional resonance [6,7]. Compared with physical vitality, which relies on the representation of physical behaviors, virtual vitality is more sensitive to the dynamic characteristics of urban space in terms of visual attraction, cultural atmosphere, and changes in public attention hotspots [8,9], providing a new perspective for understanding attention shifts in urban space.

Against the backdrop of rapid urbanization in China, population, land development expansion, and public resource allocation have long exhibited significant spatial imbalance [10]. The spatial differentiation represented by the “Heihe–Tengchong Line” indicates that population agglomeration and development disparities at the national scale profoundly influence the formation of urban density, functional organization, and center–periphery structures [11]. From this perspective, the attractiveness of urban space is not determined solely by local place characteristics but is embedded within broader spatial structures and the urbanization process. With social media platforms becoming deeply involved in daily life, online behavior is continuously influencing people’s cognition, choice, and use of physical space [12]. In this context, the attractiveness of urban space is no longer determined only by its physical environmental conditions but is also closely related to the intensity of attention accumulated and maintained in virtual space. In recent years, related studies have begun to interpret this process through theoretical perspectives such as platform urbanism and the urban attention economy [13,14]. These theories indicate that digital platforms not only record urban activities but also enable different spaces to gain online attention through recommendation and dissemination [15]. When such online attention remains at a consistently high level, it further influences offline visitation intentions and the use of physical space [16,17]. Therefore, identifying the spatial areas where virtual vitality maintains a stable high level and revealing their relationship with the structure of physical vitality space are of great significance for understanding the formation mechanism of attractiveness in contemporary urban space and promoting the transformation of virtual vitality into physical space.

In recent years, research on virtual vitality has mainly focused on virtual–physical relationships, influence mechanisms, and dynamic change, achieving some progress, yet several limitations remain:

In studies on virtual–physical relationships, existing research has used multi-source spatio-temporal big data, such as restaurant reviews [12,18], check-ins [19,20], short videos and image notes [21,22], to characterize virtual vitality, and has further incorporated POI data [23,24], mobile signaling data [25,26], and location-based population data (LBS) [27] to represent physical vitality, thereby comparing their spatial distribution differences and associations. Relevant studies have found that, in terms of spatial distribution, virtual vitality exhibits high clustering, strong centrality, and continuous diffusion, whereas physical vitality shows a relatively dispersed, multi-centered pattern constrained by physical spatial structures [28]. Studies have shown that the two types of vitality usually have a high degree of overlap in core commercial areas, but in areas with less physical pedestrian flow or weaker functional support, hotspot mismatch with “high virtual vitality and low physical vitality” also often occurs [29]. Overall, existing studies have recognized the differences in spatial patterns between the two, but the more complex interactions and supporting mechanisms between virtual vitality and physical spatial structures remain insufficiently explored.

In terms of influence mechanisms, existing studies have mainly started from built environment factors and used methods such as multiscale geographically weighted regression [30] and XGBoost [31] to identify the driving effects of factors such as POI density, functional mix, commercial facility coverage, traffic accessibility, and road connectivity on virtual vitality. However, the above findings mainly answer the question of “how virtual vitality is generated” but still lack a sufficient explanation for “why virtual vitality can be continuously maintained”. At the same time, in the representation of influencing factors, existing studies have mostly simplified physical vitality space into single indicators such as POI density or functional diversity [31]. Although this can reflect the agglomeration characteristics of physical activities to a certain extent, it is difficult to understand physical vitality space as an overall structure jointly composed of functional synergy, differences in functional composition, and locational relationships. Therefore, it is also impossible to further distinguish the differentiated effects of different physical vitality space structures on the formation and sustained maintenance of virtual vitality.

In terms of dynamic change, existing studies have shown that virtual vitality is more sensitive to events and can fluctuate many times in a short period [32], indicating that virtual vitality has a relatively complex dynamic change process. Some existing studies have begun to extend their focus to the temporal dimension, mainly from two time scales. One type takes “weekdays–weekends” as the division and describes the spatial distribution of virtual vitality under the two temporal contexts by separately calculating the average level of virtual vitality on weekdays and weekends [27,33]. The other type selects a longer time scale and focuses on the migration of virtual vitality hotspots and their spatial differences [34,35]. These studies have made some progress in characterizing the dynamic change of virtual vitality; however, these analyses mainly compare static cross-sections of virtual vitality at different time points, lack characterization of the dynamic change characteristics of virtual vitality over continuous time periods, and also make it difficult to identify the stable high-vitality spatial pattern across different temporal stages.

Based on the above research status and limitations, this study starts from three aspects—the dynamic change characteristics of virtual vitality, the identification of high-vitality stable zones, and the analysis of the virtual–physical interaction—and systematically explores the spatio-temporal dynamics of virtual vitality and its influence mechanism with the physical spatial structure. Specifically, this study focuses on the following three research questions:

(1)

What dynamic change process does virtual vitality present over continuous temporal scales?

(2)

Between different temporal stages, what transformation relationships does virtual vitality present in space, and how can high-vitality stable zones and other dynamic zoning be identified?

(3)

How do the multidimensional structural characteristics of physical vitality spaces influence the formation of high-vitality stable zones?

For this reason, this study constructs a multi-level analytical framework, which mainly includes three parts. First, this study takes the dynamic change characteristics of virtual vitality as the core and comprehensively characterizes virtual vitality from temporal variation, seasonal stability, and semantic content. Among them, based on weekly-scale data, the weekly difference method is used to identify the processes of increase, decrease, and maintenance of virtual vitality in continuous weekly sequences, thereby revealing its dynamic change characteristics; at the same time, supplemented by the comparison of stability and fluctuation characteristics at the seasonal scale, as well as TF semantic identification, supplementary analysis is conducted on the temporal variation and content expression of virtual vitality. Second, bivariate Local Moran’s I is used to analyze the hotspot results of summer and winter, identify the spatial association pattern of virtual vitality on a longer time scale, further describe its stable, sustained, or transitional states, form spatial zones with dynamic meaning, and identify high-vitality stable zones. Finally, a feature system is constructed by taking the multi-dimensional structural characteristics of physical vitality space as explanatory variables, from aspects such as functional synergy, differences in functional composition, and distance to locational centers, and the nonlinear relationship between these variables and the formation of high-vitality stable zones is further revealed.

Compared with existing studies, the main contributions of this study are threefold:

(1)

Characterizing the dynamic change characteristics of virtual vitality from continuous temporal processes, addressing the limitations of static comparison methods.

(2)

Identifying and defining high-vitality stable zones of virtual vitality, extending virtual vitality research from the perspective of spatial transformation.

(3)

Revealing the differentiated and nonlinear effects of the multidimensional structural characteristics of physical vitality spaces on the formation of high-vitality stable zones.

In summary, this study combines the dynamic process and state identification of virtual vitality with the analysis of physical spatial structure and establishes an analytical framework integrating temporal and spatial dimensions. Providing a new explanatory path for understanding the formation of urban spatial attraction in the digital era and the transformation of virtual vitality to physical space.

2. Research Materials

2.1. Description of Study Area

As the core city of Northeast China’s old industrial base, Harbin differs from some other industrial cities in the region that continue to rely heavily on traditional industrial development paths. In recent years, Harbin has gradually demonstrated a more diverse development trend [36]. Emerging industries represented by cultural tourism and modern service industries are developing at an accelerated pace and playing an increasingly important role in the transformation of cities [37]. Against this backdrop, Harbin’s online attention has shown a rapid upward trend, shifting from relatively low to high levels (

Table 1. Harbin online attention trends chart.

Year	Data Source	Ranking Name	Ranking	Feature Description
2020	21st Century Business Herald & 21 Finance App	China’s Trending Economy: 2020 Top 100 Influential Cities Ranking	33	In the early stages of online popularity, attention levels remain relatively low.
2021	21st Century Business Herald & 21 Finance App	China’s Trendy Economy: 2021 Top 100 Influential Cities Ranking	31	Online attention is limited, with relatively stable engagement levels.
2022	2022 National Tourism City Brand Influence Report	2022 Top 100 Cities for Tourism	60	The indicators lean toward tourism brand communication, resulting in a relative decline in rankings.
2023	Ouwei Data	2023 China’s Top 20 Influential Cities Index	13	Online popularity has surged significantly, ranking among the top 20 nationwide.
2024	Ouwei Data	2024 China’s Most Influential Cities Ranking	9	Its popularity continues to rise, ranking among the top 10 nationwide.
2025	The Beijing News	2025 Long-Term Popularity Ranking of Viral Cities	4	Online influence continues to grow, with popularity surging.

).

This process reflects the rapid aggregation characteristics of cities in terms of network communication power, digital narrative ability and online scene shaping, providing a representative sample for studying the dynamic changes of virtual space vitality. More importantly, Harbin shows completely different directions and volumes of online attention in summer and winter [37]. This provides a necessary comparative basis for studying the dynamic transformation of virtual vitality. In summer, online attention mainly focuses on festival activities (such as beer festivals) and natural landscape experiences, while in winter it mainly focuses on ice and snow culture. In recent years, Harbin has gradually formed a winter tourism brand represented by the Harbin International Ice and Snow Festival, relying on large-scale ice and snow theme scenic areas such as Ice and Snow World, constructing a winter urban image with highly scenarized characteristics. During the 2024–2025 ice and snow season, Harbin received 90.357 million tourists, with a total revenue of 137.22 billion yuan, forming a highly concentrated human flow and online attention in a short period, according to the latest data released by the Heilongjiang Provincial Government in 2025. These ice and snow landscapes attract many domestic and international tourists and form concentrated attention hotspots on social media, providing a seasonal comparative context for the dynamic change of virtual vitality.

Xiangfang District, Daoli District, Daowai District, Nangang District, and Songbei District (with a total area of 2342.04 km²) are areas with relatively concentrated populations and urban activities in Harbin, and are also important spatial carriers of online content production and offline activities. Therefore, this study selects these five urban districts as the study area (Figure 1).

Urban studies at similar spatial scales indicate that different grid scales significantly affect the identification results of urban spatial characteristics. Overly fine grids (e.g., 250 m) easily cause data fragmentation and reduce analysis stability, while overly coarse grids (e.g., 1000 m) may obscure local spatial characteristics; therefore, the 500 m grid is often considered a relatively balanced scale between spatial detail expression and overall pattern identification [38,39]. Based on the above research conclusions, this study adopts a 500 m × 500 m grid to divide the study area, forming a total of 9847 grids, and uses them as the basic units for subsequent analysis.

2.2. Data Sources

2.2.1. Basic Geographic Data

The basic geographic data used in this study consist of POI data and administrative boundary data. POI data were obtained from AutoNavi Maps (Gaode Map) through API access provided by the AutoNavi Open Platform, accessed in 2025. A total of 78,520 raw records were collected within the study area, covering five categories of functional places: food and beverage services, scenic attractions, shopping services, science, education and culture, and sports and leisure facilities. Referring to the Planning Standards for Urban Public Service Facilities (GB 50442–2018 [40]), semantic consolidation and functional adjustments were applied to selected subcategories to better reflect the spatial distribution characteristics of urban living-related functions. After data validation, screening, and the removal of duplicate points, 71,834 valid POI locations were retained, including 988 scenic attractions, 25,649 food and beverage establishments, 39,580 shopping facilities, 768 cultural facilities, and 4849 leisure and entertainment sites. Administrative boundary data were obtained from Tianditu, based on administrative division datasets released by the National Center for Basic Geographic Information, accessed in 2025.

2.2.2. Online Open-Source Data

Xindou is a third-party platform that provides data analytics services for the TikTok platform [23]. In this study, online check-in-related video data for Harbin were obtained through the Xindou platform.

Before obtaining the relevant data, it is first necessary to determine search keywords that match the research topic. Searching with “Harbin” as the keyword can retrieve a large amount of city-related short video content, but these videos contain many product contents named after the city, such as “Harbin Beer” and “Harbin Red Sausage,” which cannot reflect spatial use behavior. To avoid introducing noise, this study selects “Harbin check-in” as the search keyword, which is more closely related to the research. It should be noted that data obtained using a single retrieval keyword tends to reflect high-attention spaces with “check-in” attributes (such as scenic areas, shopping malls, and internet-famous restaurants), and are relatively insufficient in representing everyday spaces, thus having certain limitations in spatial representation, and the related limitations will be further analyzed in the discussion section.

At the same time, considering data stationarity and seasonal representativeness, 1 July to 31 August 2024 was selected as the summer collection period, and 1 December 2024 to 31 January 2025 was selected as the winter collection period, obtaining 3244 summer-related videos and 6296 winter-related videos.

Aiming at the problems such as ambiguous addresses, misaligned classification and name changes in the location information associated with videos, a unified data filtering rule is adopted for processing in the research:

(1)

Locations without clear addresses or accurate spatial positioning were removed.

(2)

Locations with seasonal name changes but unchanged business types were merged.

After data cleaning, 529 valid geographic coordinates were obtained from summer videos and 896 from winter videos. These coordinates constitute the social media check-in points used in subsequent spatial analyses. The video data also include user-posted text information, which provides the basis for semantic feature analysis of virtual vitality.

2.2.3. Functional Classification

Considering that urban activities mainly revolve around consumption, staying, and experience, and that social media check-in behavior also depends on specific offline place types, this study, under a unified analysis framework of virtual and physical vitality, classifies both physical POI points and TikTok check-in points into five categories: tourist attractions, cuisine, shopping, leisure and entertainment, and culture. This classification reflects differences among space types in activity attraction, consumption scenarios, and leisure experience, and meets the requirements of the corresponding analysis of virtual–physical vitality spaces, providing a basis for identifying the influence of physical vitality spatial structure on virtual vitality. In the specific classification process, the subcategories of physical POIs are mainly based on the original classification of Gaode Map, with semantic merging and functional adjustment referring to the Urban Public Service Facilities Planning Standard (GB 50442—2018); the subcategories of TikTok check-in points refer to the platform’s original categories and are integrated accordingly based on research needs. The specific classification is shown in

Table 2. Functional categories of virtual and physical vitality spaces.

Type	TikTok Check-In POI	Gaode Maps POI
Tourist attractions	(6) Botanical garden, aquarium, user-generated attractions, scenic spots, zoo, tourist attractions	(7) Religious sites, tourist attractions, botanical gardens, zoos, memorial halls, parks and plazas, scenic spots
Cuisine	(20) Buffet, Cantonese cuisine, beverages, music restaurant, western cuisine, barbecue, Japanese cuisine, theme restaurant, bread and pastries, fast food and snacks, grilled meat, hot pot, Korean cuisine, seafood, Southeast Asian cuisine, Northeastern Chinese cuisine, skewered hot pot, Sichuan cuisine, Shanghai-style Jiangsu–Zhejiang cuisine, Russian cuisine	(10) Chinese restaurants, casual dining establishments, ice cream shops, coffee shops, foreign restaurants, fast-food restaurants, dessert shops, bakeries, tea houses, food and beverage-related venues
Shopping	(5) Specialty stores, commercial streets, retail stores and supermarkets, department stores, shopping malls	(10) Specialty stores, convenience stores, shopping malls, retail centers, electronics and home appliances stores, pawnshops, auction houses, commercial streets, markets, department stores and supermarkets
Leisure and entertainment	(14) Fitness centers, entertainment venues, amusement parks, bathhouses and massage parlors, internet cafes, board and card game establishments, activity centers, escape rooms, photography and travel photography services, bars, vacation resorts and spas, movie theaters, fruit-picking farms, parks and plazas	(17) Recreational venues, fishing parks, water activities, amusement parks, fruit-picking gardens, campgrounds, karaoke and dance entertainment, nightclubs, card and board game rooms, internet cafes, arcades, bars, health resorts and vacation spots, theaters, concert halls, sports stadiums, sports service facilities
Culture	(7) Exhibition halls, concert halls, libraries, art galleries, theaters, museums, cultural and creative centers	(10) Convention and exhibition center, museum, library, planetarium, exhibition hall, cultural palace, performing arts troupe, archives, science and technology museum, art museum

.

3. Methods

We propose an analysis framework based on TikTok video data to examine the temporal evolution, spatial transition, and the association mechanisms between dynamic zoning of virtual vitality and physical vitality spaces (as shown in Figure 2).

3.1. Measurement of Virtual Vitality

Based on the construction of virtual vitality measurement indicators, this section comprehensively characterizes virtual vitality from three aspects: temporal variation characteristics, seasonal-scale characteristics, and semantic expression.

3.1.1. Measurement Indicators and Calculation Methods of Virtual Vitality

The spatial manifestation of virtual vitality is reflected in the degree of activity aggregation at different locations, and therefore, its spatial intensity needs to be quantitatively represented. Existing studies indicate that social media data can serve as important proxies for characterizing virtual vitality or the spatial distribution of urban activities [31,41]. Based on this, kernel density estimation (kernel density estimation, KDE) is used to make spatially continuous the TikTok check-in data, and the results are taken as the virtual vitality intensity indicator $\mathit{VP}$ to characterize the aggregation of virtual activities in urban space.

It should be noted that, considering the total number of points, the extent of the study area, and the 500 m grid scale, the existing TikTok check-in points show a certain degree of spatial sparsity, which may affect the results of kernel density estimation. Therefore, this study does not interpret the kernel density results as a precise estimation for a single grid unit but treats them as a representational tool for identifying the overall spatial pattern and spatial aggregation characteristics of virtual vitality. Existing studies have shown that under conditions where the point scale and analysis scale are similar, kernel density estimation can still support the identification of overall spatial distribution patterns [38,42].

In the selection of weights, to simultaneously reflect the spatial occurrence intensity and information dissemination characteristics of virtual activities, the number of location occurrences and the corresponding video likes are used as weight variables. Existing studies, when characterizing virtual vitality, adopt a weighting approach combining spatial behavior intensity and information dissemination intensity, and apply a weighting ratio of 7:3 [23,29]. This study follows this approach, taking the number of location occurrences as the indicator of spatial occurrence intensity, taking the number of likes as the indicator of information dissemination characteristics, and applying the 7:3 weighting setting for calculation. To verify the rationality of the weighting scheme, 6:4 and 8:2 are selected as comparison scenarios for sensitivity analysis. Taking high-value areas (the top 10% of spatial units) as the benchmark, two weeks are selected in both summer and winter, and the spatial distribution results under different weight settings are compared comprehensively. The results show that the spatial overlap of high-value areas under different weight settings is generally high: the overlap between the 7:3 and 8:2 weight settings ranges from 0.984 to 0.9528, with an average of 0.9743; the overlap between the 7:3 and 6:4 weight settings ranges from 0.992 to 0.9811, with an average of 0.9852 (see

Table 3. Comparison of spatial overlap of high-value areas under different weight settings.

Comparison Weight Scheme	Max	Min	Mean
8:2	0.984	0.9528	0.9743
6:4	0.992	0.9811	0.9852

for details). The results indicate that different weight settings have limited impact on the spatial distribution of high-value areas, and the weighting setting shows strong robustness and is not sensitive to parameter changes. In addition, to reduce the influence of scale differences between variables, both variables are transformed using ln(x + 1) before weight aggregation. The calculation formula is as follows:

$$w_i=0.7·\ln ({\mathit{Count}}_i+1)+0.3·\ln ({\mathit{Like}}_i+1)$$

(1)

Here, $w_i$ denotes the synthesized weight of grid $i$, ${\mathit{Count}}_i$ represents the occurrence frequency, and ${\mathit{Like}}_i$ indicates the number of likes associated with grid $i$.

3.1.2. Weekly-Scale Dynamic Change Characterization and Change Category Identification

After obtaining the spatial distribution of virtual vitality intensity, to further characterize its change process in the temporal dimension, a difference method between adjacent time periods is used to dynamically measure virtual vitality. Comparisons of multiple static cross-sections cannot reveal the dynamic change of virtual vitality itself. This study represents the dynamic change process of virtual vitality through differences between adjacent time periods. The study uses natural weeks as the temporal analysis units (the selected weeks and their time ranges are shown in

Table 4. Natural week division for summer and winter virtual vitality analysis.

Weekly Serial Number	Summer Time Period (2024)	Winter Time Period (2024–2025)
Week 1	1 Jul (Mon)–7 Jul (Sun)	2 Dec (Mon)–8 Dec (Sun)
Week 2	8 Jul (Mon)–14 Jul (Sun)	9 Dec (Mon)–15 Dec (Sun)
Week 3	15 Jul (Mon)–21 Jul (Sun)	16 Dec (Mon)–22 Dec (Sun)
Week 4	22 Jul (Mon)–28 Jul (Sun)	23 Dec (Mon)–29 Dec (Sun)
Week 5	29 Jul (Mon)–4 Aug (Sun)	30 Dec (Mon)–5 Jan (Sun)
Week 6	5 Aug (Mon)–11 Aug (Sun)	6 Jan (Mon)–12 Jan (Sun)
Week 7	12 Aug (Mon)–18 Aug (Sun)	13 Jan (Mon)–19 Jan (Sun)
Week 8	19 Aug (Mon)–25 Aug (Sun)	20 Jan (Mon)–26 Jan (Sun)

).

And the weekly-scale virtual vitality change value $D_t\left(p\right)$ for each spatial unit is obtained. This value is used to represent the enhancement, decline, and relative stability of virtual vitality between consecutive weeks. The calculation formula is as follows:

$$D_t\left(p\right)=V_{t+1}\left(p\right)-V_t\left(p\right)$$

(2)

For change level classification, because some differenced values show asymmetric and skewed distributions, the absolute values on the $D_t\left(p\right)$ > 0 and $D_t\left(p\right)$ < 0 sides are separately classified into two levels using the natural breaks (Jenks) method, extracting the strong–weak thresholds $\mathit{TP}$ and $\mathit{TN}$. At the same time, the break point closest to zero is extracted from the overall differenced distribution as $Z0$ to represent the low-level stable interval. To avoid the influence of extreme values on one side of the distribution, the unified threshold is set as the minimum of $\mathit{TP}$, $\mathit{TN}$, and $Z0$, that is:

$$T_0=\min (T_P,T_N,Z_0)$$

(3)

Based on this, a five-level change classification system is constructed: “significant decrease–slight decrease–dynamic balance–slight increase–significant increase.”

$$D_t(p)=\{\begin{array}{ll}\mathrm{Significant} \mathrm{Decrease},& D_t\left(p\right)<-T_N\\ \mathrm{Slight} \mathrm{Decrease},& -T_N\le D_t\left(p\right)<-T_0\\ \mathrm{Dynamic} \mathrm{Balance},& -T_0\le D_t\left(p\right)\le T_0\\ \mathrm{Slight} \mathrm{Increase},& T_0<D_t\left(p\right)\le T_P\\ \mathrm{Significant} \mathrm{Increase},& D_t\left(p\right)>T_P\end{array}$$

(4)

3.1.3. Seasonal Mean and Fluctuation Levels of Virtual Vitality

Based on the characterization of the dynamic change of virtual vitality, it is further measured from the perspectives of the overall level and stability. By averaging the kernel density values of all weeks within the season for each spatial unit in the study area, the seasonal mean of virtual vitality is obtained. Compared with single-week data, this approach reduces the influence of short-term fluctuations, thereby reflecting the overall intensity of virtual vitality over the entire season and enabling comparison of virtual vitality intensity between different seasons; the calculation formula is as follows:

$${\displaystyle \frac{1}{N\times W}}\sum _{i=1}^N\sum _{t=1}^{W}K{D}_{i,t}$$

(5)

Here, $K{D}_{i,t}$ denotes the kernel density value of grid $i$ in week $t$, $N$ is the total number of spatial units, and $W$ is the total number of weeks.

Based on the standard deviation concept, a Seasonal Variability Index ($\mathit{SVI}$) is constructed. In this study, this index is used to reflect the fluctuation amplitude of virtual vitality. A larger value indicates stronger intra-seasonal variation and weaker stability, while a smaller value indicates smaller weekly-scale changes and relatively stable seasonal patterns.

$$\begin{array}{c}\mathit{SVI}={\displaystyle \frac{1}{M}}{\displaystyle \sum _{p=1}^M}\sqrt{{\displaystyle \frac{1}{N}}{\displaystyle \sum _{t=1}^N}{\left(V_t\left(p\right)-{\displaystyle \overline{V}}\left(p\right)\right)}^2}\end{array}$$

(6)

Here, $p$ denotes a spatial unit, $\mathit{Vt}\left(p\right)$ is the virtual vitality of unit $p$ in week $t$, $\overline{V}\left(p\right)$ represents the seasonal weekly mean for unit $p$, $M$ is the total number of spatial units, and N denotes the number of natural weeks in the season.

3.1.4. TF–IDF-Based Extraction of Semantic Features of Virtual Vitality

TF–IDF is a text analysis method used to measure the relative importance of words. By combining term frequency and cross-text distinctiveness, it identifies keywords with strong semantic relevance [43].

In this study, the “video descriptions” in the video data are used as the objects of semantic analysis, and the text corpora for summer and winter are tokenized. Based on tidytext, TF–IDF values are calculated for each term, and the top 15 keywords for summer and winter are selected in descending order. This identifies high-frequency semantic features and differences in virtual vitality between summer and winter, providing a basis for understanding seasonal variation in virtual vitality.

3.2. Spatial Analysis Methods of Virtual Vitality

3.2.1. Hotspot Analysis

Hot spot analysis can identify spatial clusters of high and low virtual vitality based on the distribution of attribute values weighted by spatial neighborhoods [44,45]. In this study, the Getis–Ord statistic is used to identify hot and cold spots in summer and winter, thereby capturing the spatial clustering characteristics of virtual vitality.

As different analysis methods serve different research purposes, the selection of unit observation values should be distinguished accordingly. Specifically, Section 3.1 measures virtual vitality based on kernel density estimation, aiming to comprehensively characterize its spatial occurrence intensity and information dissemination characteristics; therefore, the weighted result of the number of location occurrences and video likes is selected as the indicator of virtual vitality. In this section, the hot spot analysis is mainly used to identify the spatial aggregation pattern of virtual vitality, with greater emphasis on actual spatial occurrence frequency; therefore, the number of check-in points is selected as the observation variable. The calculation formula is as follows:

$$G_i^{\ast }={\displaystyle \frac{\sum _j{w}_{\mathit{ij}}{x}_j-\overline{X}\sum _j{w}_{\mathit{ij}}}{S\sqrt{\frac{n\sum _j{w}_{\mathit{ij}}^2-{\left(\sum _j{w}_{\mathit{ij}}\right)}^2}{n-1}}}}$$

(7)

Here, $x_j$ is the observed value of spatial unit $j$ (defined as the number of check-in points within each grid cell), $w_{\mathit{ij}}$ denotes the spatial weight between units i and $j$, $\overline{X}$ and $S$ are the mean and standard deviation of all observed values, respectively, and $n$ is the total number of spatial units.

3.2.2. Bivariate Local Spatial Autocorrelation Analysis

To further reveal local association characteristics between the spatial patterns of virtual vitality in summer and winter, this study uses bivariate local spatial autocorrelation analysis. In the analysis, the standardized $G_i^{\ast }$ statistic of summer virtual vitality is used as variable A, and the standardized $G_i^{\ast }$ statistic of winter virtual vitality is used as variable B, to identify their local spatial correlation. The specific formula is as follows:

$$I_p=z(A_p)\sum _q{w}_{\mathit{pq}}\mathrm{z}(B_q)$$

(8)

where $w_{\mathit{pq}}$ denotes the spatial weight matrix constructed based on spatial adjacency relationships using the first-order Queen contiguity criterion, and $z(·)$ represents the standardization transformation. The significance level was set at $\alpha =0.01$, and statistical significance was identified through Monte Carlo permutation tests, resulting in five types of spatial units: High–High (HH), Low–High (LH), High–Low (HL), Low–Low (LL), and non-significant areas.

Based on the local spatial association structure and significance results, the seasonal dynamic spatial pattern of virtual vitality is divided into four statistically significant region types: high-vitality stable zones (HH), where both summer and winter show significant high-value clustering; vitality-declining zones (HL), where summer shows significant high values but weakens in winter; vitality-enhancing zones (LH), where summer is weaker and winter shows significant enhancement; and low-vitality stable zones (LL), where both summer and winter show significant low-value clustering.

3.3. Mechanism Modeling

3.3.1. Indicator Screening and Feature Construction

By integrating existing studies on virtual vitality and urban vitality, this study constructs an explanatory variable system for the stable maintenance of high virtual vitality from two sources: common influencing factors widely discussed in previous virtual vitality studies, and supplementary factors introduced from urban vitality studies to better capture spatial structural characteristics. The former mainly includes three types of variables: distance to central areas [46,47], physical POI density [48,49], and functional synergy [50,51]. The latter includes four additional aspects: functional composition dissimilarity [52], neighborhood relative density [53], center–periphery gradient [54], and functional dominance [55]. These supplementary aspects are operationalized as functional composition dissimilarity indicators, neighborhood relative density indicators, ring-based spatial gradient indicators, and functional dominance indicators. Together with the three commonly used factors, they form an explanatory variable system comprising seven categories of indicators. Before modeling, variance inflation factor (VIF) testing was conducted on the candidate variables [56], and variables with VIF > 5 were removed. Finally, 16 features were retained for model training. The meaning, variable composition, and calculation method of each indicator are presented in

Table 5. Final set of features used for model training.

Category	Indicators	Definition	Interpretation
Centrality-based location indicator	Distance to activity center (tourism)	Euclidean distance from a grid cell to the weighted centroid of a given POI category (m)	Measure the spatial distance between a grid cell and the nearest activity center, reflecting accessibility to core functional nodes
POI count indicators	Catering facility count, leisure facility count, cultural facility count, shopping facility count, and tourist attraction count	POI count within a grid cell (25 ha)	Quantify the functional scale and service capacity of a node by counting the number of POIs of specific categories within a grid
Functional dominance indicators	Functional dominance	Ratio of the count of the most frequent POI category to the total POI count within a grid cell (%)	Measure whether a single function dominates within a grid by comparing the proportion of one POI category to the total number of POIs
Functional composition dissimilarity indicators	Functional composition dissimilarity (culture), functional composition dissimilarity (shopping)	Using the total number of a given POI category in the central grid cell and all grid cells within its 750 m neighborhood as the denominator, the absolute difference between the proportion of that POI category in the central grid cell and the proportion in the neighboring grid cells (unitless)	Quantify the compositional difference of a specific function between a grid cell and its surrounding 3 × 3 neighborhood, describing the inconsistency in functional distribution between the cell and its surrounding area
Neighborhood relative density indicators	Tourism density ratio, cultural density ratio	Ratio of the number of a given POI category within a grid cell to the average number of the same POI category within its 750 m neighborhood (unitless)	Evaluate the degree of local concentration of POIs within a grid relative to its surrounding area
Ring-based spatial gradient indicators	Shopping ring gradient, leisure ring gradient, tourist ring gradient	Calculation of the ratio of the difference in the number of a given POI category between the outer ring (750–1250 m) and the core ring (0–750 m) centered on a grid cell to the total number of the two rings (unitless)	Capture spatial variation in POI density between inner and outer buffers, reflecting core–periphery differentiation
Functional synergy indicators	Functional synergy (culture–leisure), functional synergy (tourism–leisure)	Calculation of the ratio of the square root of the product of the counts of two POI categories within the same grid cell to the total POI count within that grid cell (unitless)	Measure the degree of co-location and interaction between different functional categories within a grid

. These indicators involve locational conditions, functional scale and aggregation characteristics, functional structure relationships, and spatial structure differences, and can characterize the physical vitality spatial structure from multiple dimensions.

3.3.2. Random Forest

Compared with traditional regression models that focus on significance testing, or other tree-based algorithms (such as GBDT, XGBoost, and AdaBoost), the Random Forest model is more suitable for revealing the differentiated effects of different factors on the target variable across different ranges, and is therefore often used in explanatory studies of influence mechanisms [57]. This is consistent with the research focus of this study; therefore, the Random Forest model is selected to analyze the relationships and response characteristics between different physical spatial features and target spatial types. To enhance the rationality of model selection, this study further introduces the LightGBM model as a comparison model and compares different models based on classification performance and result interpretation requirements.

In the setting of the dependent variable, high-vitality stable zones (HH zones), characterized by high virtual vitality and inter-seasonal stability, are used to represent spatial types with stable high vitality; therefore, this type of space is taken as the dependent variable. It should be noted that the “stable maintenance of high virtual vitality” mentioned here is an extended expression based on the spatial classification of HH zones, representing a conceptual generalization of this spatial phenomenon, rather than a direct modeling of the temporal persistence process of virtual vitality. The independent variables include the 16 physical vitality spatial structure indicators mentioned above.

The Random Forest model is implemented in R, mainly using the caret and ranger packages. Samples are split into training and testing sets at a 7:3 ratio, and model parameters are tuned on the training set using five-fold cross-validation repeated three times. For parameters, key parameters such as mtry and min.node.size are searched within candidate grids.

In terms of model interpretation, feature importance analysis [58] and partial dependence analysis [59] are used as the main explanatory methods. Among them, feature importance analysis is used to rank the relative contributions of explanatory variables in the model, identify key spatial structure factors affecting the formation of high-vitality stable zones, and compare differences in variable importance; partial dependence analysis is used to characterize the direction and trend of key variables on the stable maintenance of virtual vitality across different value ranges, thereby identifying nonlinear response relationships and potential threshold characteristics. Based on the above analysis, the model results are further translated into specific spatial optimization bases.

4. Results

4.1. Dynamic Features and Seasonal Differences of Virtual Vitality

4.1.1. Weekly Change-Type Identification of Virtual Vitality

Based on weekly kernel density differentials, virtual vitality in the study area exhibited five change states between consecutive weeks: slight decrease, significant decrease, dynamic balance, slight increase, and significant increase (Figure 3 and Figure 4). The area proportions of each category are shown in Figure 5, and the temporal sequences for summer and winter are listed in

Table 6. Time intervals of weekly kernel density difference analysis.

Change Phase	Summer Difference Interval (Week t–Week t − 1)	Winter Difference Interval (Week t–Week t − 1)
1	8 Jul–14 Jul → 1 Jul–7 Jul	9 Dec–15 Dec → 2 Dec–8 Dec
2	15 Jul–21 Jul → 8 Jul–14 Jul	16 Dec–22 Dec → 9 Dec–15 Dec
3	22 Jul–28 Jul → 15 Jul–21 Jul	23 Dec–29 Dec → 16 Dec–22 Dec
4	29 Jul–4 Aug → 22 Jul–28 Jul	30 Dec–5 Jan → 23 Dec–29 Dec
5	5 Aug–11 Aug → 29 Jul–4 Aug	6 Jan–12 Jan → 30 Dec–5 Jan
6	12 Aug–18 Aug → 5 Aug–11 Aug	13 Jan–19 Jan → 6 Jan–12 Jan
7	19 Aug–25 Aug → 12 Aug–18 Aug	20 Jan–26 Jan → 13 Jan–19 Jan

.

Virtual vitality in summer and winter shows two distinctly different dynamic change patterns within seasons: summer presents a pattern of “rapid increase—high-level convergence—gradual decrease,” while winter shows a pattern of “balanced start—concentrated increase—low-amplitude fluctuation.”

Specifically, in summer, the first change shows a significant leap, with growth-type areas accounting for about 30%. The slight increase and significant increase account for 23.17% and 6.70%, respectively. Spatially, growth areas expand outward with the significant increase area as the core. This change corresponds closely in time to the concentrated launch of summer cultural and tourism projects in early July, such as the opening of the Harbin Beer Festival on July 6. At the same time, significant increase areas cover seasonal outdoor sites including Sun Island and the Northeast Tiger Park, reflecting the rapid response of virtual vitality to the launch rhythm of physical activities.

This enhancement did not continue. From the second to the fourth changes, the proportion of increase-type areas gradually declined, while the proportion of dynamic balance increased significantly and remained at a high level. In the third change, the proportion of dynamic balance reached 80.68%, the peak of this stage. Overall, virtual vitality during this period shows spatial convergence of high-level fluctuations.

In the fifth change, the proportion of dynamic balance dropped to 68.33%, while the proportions of decrease-type and increase-type areas became closer, indicating intensified changes in summer virtual vitality. During this stage, the proportion of dynamic balance declined rapidly, and the focus shifted faster, with the same locations showing clearly different vitality states across adjacent weeks. Spatially, changes are mainly concentrated around landmark nodes and present a diffusion pattern that weakens outward from the core. In late August, as seasonal activities gradually ended, the proportion of dynamic balance in the seventh change declined to 57.56%, while decrease-type areas were clearly higher than increase-type areas, reflecting an end-of-season adjustment of summer tourism demand.

In the first winter change, the proportions of all categories are relatively balanced. This may be because this stage corresponds to the preheating period of the ice and snow season. As temperatures decrease and outdoor activities gradually reduce, virtual vitality does not immediately enter a unidirectional increase. In the second change, the proportion of increase-type areas rose rapidly to 31.68%, which corresponds closely to the official opening of the Ice and Snow World on 21 December, and virtual vitality showed concentrated growth.

From the third to the fourth changes, the proportion of dynamic balance remained stable at about 68%, with increase-type areas consistently higher than decrease-type areas, indicating a stable high-vitality state. With time, in the fifth change, the proportion of decrease-type areas rose to 34.07%, corresponding to the decline in visitor flow after the New Year holiday. Virtual vitality shifted from a stable high level to a downward adjustment state and entered a recession transition period. During this period, decrease-type areas were mainly distributed around core consumption areas, while local increase-type areas remained around Central Street and Saint Sophia Cathedral. In the sixth and seventh changes, both increase-type and decrease-type grids remained within a small fluctuation range, and virtual vitality entered a low-range fluctuation stage.

4.1.2. Seasonal Differences in the Mean and Fluctuation Levels of Virtual Vitality

From the perspectives of mean intensity and fluctuation characteristics, virtual vitality shows certain differences across seasons: the mean intensity is 0.1406 in winter and 0.0882 in summer; the seasonal variation intensity index is 0.2675 in winter and 0.3498 in summer.

Virtual vitality in winter shows a higher overall level and smaller intra-seasonal fluctuation, indicating stronger temporal stability; in contrast, summer shows lower overall intensity but significantly greater fluctuation, presenting a “low intensity–high fluctuation” pattern.

From the perspective of spatial distribution (Figure 6), the fluctuation range of virtual vitality in summer is more extensive. Apart from the core area, in summer, various leisure spaces experience fluctuations of different intensities of vitality. In contrast, although the virtual vitality in winter has formed a significant concentration in the core scenic spots, the overall spatial fluctuation degree is more convergent.

4.1.3. Seasonal Differences in Semantic Features of Virtual Vitality

TF-IDF analysis indicates that virtual vitality in summer and winter shows significant differences in thematic structure and expression. High-weight keywords in summer texts show a dispersed pattern and lack a clear semantic theme, and their weights decrease rapidly with ranking. Keyword content mainly focuses on temporal states and subjective expressions, such as summer vacation, summer day, and coolness, showing strong emotional and generalized characteristics. Winter keywords show a highly concentrated thematic structure with clear scenario orientation, with the top 15 high-weight keywords mainly centered on “ice and snow”-related events, directly corresponding to specific activities and spatial scenes, such as Snow Expo, Park opening, Snowfield, Ice and Snow Festival, and Icy and snowy landscape (Figure 7). At the same time, the keyword weights in winter decrease more gradually with ranking, showing greater persistence and concentration in thematic expression.

4.2. High-Vitality Stable Zones and Related Dynamic Zones

4.2.1. Seasonal Hotspot Patterns of Virtual Vitality

The hotspot analysis results for summer and winter (Figure 8) exhibit clear seasonal contrasts. Summer hot spots are mainly distributed in spaces dominated by open leisure activities and form continuously expanding clustered structures, while winter hot spots are clearly concentrated in spaces represented by commercial complexes and core attractions, showing a more concentrated and stable aggregation pattern.

Overall, summer virtual vitality hot spots have a wider distribution range and stronger spatial extension, while winter hot spots tend to concentrate on a few high-intensity core nodes.

4.2.2. Identification of High-Vitality Stable Zones and Other Dynamic Zone Types

The results of the bivariate local Moran’s I analysis indicate that seasonal virtual vitality in the study area forms four statistically significant types of local spatial associations, namely high-vitality stable zones (HH), vitality-declining zones (HL), vitality-enhancing zones (LH), and low-vitality stable zones (LL). The spatial distribution patterns and proportional shares of each type are shown in Figure 9 and

Table 7. Significant grid cell statistics for summer and winter virtual vitality using bivariate Moran’s I.

Bivariate LISA Type	Spatial Category	Number of Grids	Proportion (%)
HH	high-vitality stable zones	422	48.84
LH	vitality-enhancing zones	62	7.18
HL	vitality-declining zones	50	5.79
LL	low-vitality stable zones	330	38.19

.

Among these, the High–High (HH) zones include 422 units, accounting for 48.84% of the total. They form spatially continuous core clusters within the main urban area, indicating a high degree of spatial stability. The Low–High (LH) zones are mainly distributed around the HH zones, showing clear adjacency and encircling patterns. Although vitality levels in these zones are relatively low in summer, they increase markedly in winter in response to the intensification of adjacent high-vitality areas, forming a regular peripheral distribution pattern that reflects strong dependence on the core hotspots. By contrast, the High–Low (HL) zones comprise only 50 units, accounting for 5.79% of the total, and are relatively few in number and spatially dispersed, indicating pronounced seasonal sensitivity. In addition, the Low–Low (LL) zones comprise 330 units, accounting for 38.19% of the total. These zones show no significant clustering of virtual vitality in either summer or winter and are mainly distributed along the periphery of the study area.

4.3. Influencing Factors and Response Patterns of the Stable Maintenance of High Virtual Vitality

To verify the robustness of model selection, this study compares the Random Forest (RF) and LightGBM (LGBM) models (

Table 8. Comparison of model performance between Random Forest (RF) and LightGBM (LGBM).

Model	Dataset	AUC	Precision	Recall	F1-score	Accuracy
LGBM	Train	0.9465	0.875	0.943	0.907	0.906
LGBM	Test	0.9325	0.866	0.921	0.892	0.891
RF	Train	0.9739	0.871	0.980	0.922	0.919
RF	Test	0.9387	0.873	0.984	0.925	0.922

).

The results show that the two models have similar AUC levels on the test set, while the Random Forest performs slightly better in Precision, Recall, F1-score, and Accuracy, without obvious overfitting. Considering overall predictive performance, stability, and interpretability, the Random Forest model is selected for subsequent feature importance and partial dependence analysis.

The main indicators of the model perform well, with no overfitting observed, indicating that the model has stable explanatory ability.

4.3.1. Feature Importance

The results of the feature importance analysis (Figure 10) show that different types of features have significant differences in their explanatory power for the formation of high-vitality stable zones, among which centrality-based location indicator and functional composition dissimilarity indicators are the most critical.

Specifically, functional composition dissimilarity (culture) ranks first in importance. This indicates that, at the neighborhood scale, differences in cultural functional structure have the greatest influence on the stable maintenance of high virtual vitality. Distance to activity center (tourism) ranks second, indicating that this variable is of high importance in the Random Forest model, suggesting that the distance between spatial units and tourism activity centers has a strong correlation with the stable maintenance of high virtual vitality. The third-ranked functional composition dissimilarity (shopping) further confirms the significant influence of the functional composition dissimilarity at the neighborhood scale on the stable formation of virtual vitality. The above three characteristics together constitute the first-tier influencing factors of the stable partition of virtual vitality.

At the medium importance level, indicators such as functional dominance, catering facility count, and leisure facility count have certain explanatory power. These characteristics better reflect the local functional structure or the scale of specific business formats. It has an auxiliary role in the stable formation of virtual vitality, but it is difficult to independently dominate the spatial partitioning results. In contrast, the overall importance of neighborhood relative density indicators and ring-based spatial gradient indicators is relatively low. It indicates that simple circus decrement or local density differences have limited explanatory power for the stable partitioning of virtual vitality.

In addition, most functional synergy indicators also display relatively low importance values, suggesting that the stable formation of virtual vitality does not rely on complex functional synergies, but instead depends more on core structural features with clear spatial orientation.

4.3.2. Nonlinear Response Relationships of Key Influencing Factors

The partial dependence analysis results (Figure 11) (Table 9) show that different types of indicators exhibit clear differences in their response forms on the probability of HH zone formation, mainly presenting nonlinear relationships such as threshold type, unimodal type, and U-shaped type. Overall, the effects of most key variables do not continuously increase with value but are mainly concentrated within specific ranges.

(1)

Centrality-Based Location Indicators

Distance to activity center (tourism) shows a threshold-type decreasing relationship with the probability of HH zone formation. Before reaching the turning point, increasing distance leads to a rapid decrease in the probability of HH zone formation. After the distance reaches the threshold, the probability change becomes more gradual. This indicates that the effect of this feature on HH zone formation is mainly concentrated in areas close to the activity core.

(2)

POI Count Indicators

These indicators mainly show threshold-type and unimodal responses, but differences exist across functional types. Among them, catering facility count and shopping facility count show threshold-type increasing relationships, with turning points at 11.88 and 21.80, respectively. This indicates that within the low-to-middle range, the probability of HH zone formation increases rapidly with higher POI counts, after which the growth becomes more gradual. By contrast, tourist attraction count shows a unimodal response, reaching a maximum at 11.63, after which further increases lead to a decrease in the probability of HH zone formation. This indicates that the response pattern of tourist attraction count differs from other POI count indicators. Within a zone, a higher number of tourist attraction POIs is not always better, and continuous increases can reduce the probability of HH zone formation.

(3)

Functional Dominance Indicators

Functional dominance indicators show a unimodal response. Its turning point is at x = 0.14, after which the probability of HH zone formation increases rapidly and reaches a maximum at x = 0.43. After that, as the share of the dominant function further increases, the probability of HH zone formation gradually decreases.

(4)

Functional Composition Dissimilarity Indicators

These indicators exhibit different nonlinear patterns in the PDP. Functional composition dissimilarity (culture) shows a threshold-type relationship, with a turning point at x = 0.22. Before the threshold, increasing dissimilarity raises the probability of HH zone formation. After the threshold is exceeded, the increase becomes more gradual. By contrast, functional composition dissimilarity (shopping) shows a unimodal response and reaches its maximum at x = 0.88, indicating that there is an optimal level of dissimilarity in shopping-related functional spaces, while both too low and too high levels are unfavorable for HH zone formation.

(5)

Other Structural Indicators

By contrast, neighborhood relative density indicators, ring-based spatial gradient indicators, and functional synergy indicators, which rank lower in feature importance, exhibit threshold-type, unimodal-type, or U-shaped relationships, all showing clear interval dependence.

Figure 11. Partial dependence response relationships for different characteristics.

Figure 11. Partial dependence response relationships for different characteristics.

Table 9. Partial dependence results: key peaks, inflection points, and response shapes.

Feature	Peak Value	Peak Response	Inflection Point	Inflection Response	Shape
Functional composition dissimilarity (shopping)	0.88	0.58	0.04	0.44	Unimodal
Functional composition dissimilarity (culture)	1.00	0.66	0.22	0.54	Threshold-type
Catering facility count	190.04	0.56	11.88	0.51	Threshold-type
Distance to activity center (tourism)	0.01	0.58	0.27	0.45	Threshold-type
Functional dominance	0.43	0.55	0.14	0.47	Unimodal
Shopping facility count	392.33	0.54	21.80	0.50	Threshold-type
Shopping ring gradient	0.22	0.50	−0.51	0.48	Unimodal
Tourist attraction count	11.63	0.54	4.29	0.52	Unimodal
Tourism density ratio	38.37	0.53	1.63	0.52	Threshold-type
Tourist ring gradient	0.84	0.50	−0.02	0.49	U-shaped
Functional synergy (tourism–leisure)	0.37	0.55	0.06	0.51	Threshold-type
Functional synergy (culture–leisure)	0.16	0.53	0.02	0.51	Threshold-type
Cultural facility count	4.41	0.55	1.10	0.52	Threshold-type
Cultural density ratio	4.41	0.52	0.80	0.50	Unimodal
Leisure facility count	108.39	0.56	4.61	0.52	Threshold-type
Leisure ring gradient	−0.14	0.50	0.02	0.49	Unimodal

Note: The distance to activity center (tourism) indicator was normalized before Random Forest modeling; therefore, the peak and inflection values reported in this table do not have actual physical meaning.

Table 9. Partial dependence results: key peaks, inflection points, and response shapes.

Feature	Peak Value	Peak Response	Inflection Point	Inflection Response	Shape
Functional composition dissimilarity (shopping)	0.88	0.58	0.04	0.44	Unimodal
Functional composition dissimilarity (culture)	1.00	0.66	0.22	0.54	Threshold-type
Catering facility count	190.04	0.56	11.88	0.51	Threshold-type
Distance to activity center (tourism)	0.01	0.58	0.27	0.45	Threshold-type
Functional dominance	0.43	0.55	0.14	0.47	Unimodal
Shopping facility count	392.33	0.54	21.80	0.50	Threshold-type
Shopping ring gradient	0.22	0.50	−0.51	0.48	Unimodal
Tourist attraction count	11.63	0.54	4.29	0.52	Unimodal
Tourism density ratio	38.37	0.53	1.63	0.52	Threshold-type
Tourist ring gradient	0.84	0.50	−0.02	0.49	U-shaped
Functional synergy (tourism–leisure)	0.37	0.55	0.06	0.51	Threshold-type
Functional synergy (culture–leisure)	0.16	0.53	0.02	0.51	Threshold-type
Cultural facility count	4.41	0.55	1.10	0.52	Threshold-type
Cultural density ratio	4.41	0.52	0.80	0.50	Unimodal
Leisure facility count	108.39	0.56	4.61	0.52	Threshold-type
Leisure ring gradient	−0.14	0.50	0.02	0.49	Unimodal

Note: The distance to activity center (tourism) indicator was normalized before Random Forest modeling; therefore, the peak and inflection values reported in this table do not have actual physical meaning.

5. Discussion

5.1. Dynamic Change Patterns of Virtual Vitality and Their Seasonal Responsiveness

The key change points of virtual vitality in the five urban districts of Harbin in summer and winter were basically synchronized with physical events such as the opening ceremony of the Beer Festival and the opening of Ice and Snow World, indicating that virtual vitality is not a random fluctuation but a process shaped by both temporal rhythms and spatial scenes [60,61]. Different from existing studies, which mainly summarize the dynamic characteristics of virtual vitality through differences in vitality intensity across different time sections, this study uses the weekly difference method to describe the complete dynamic evolution process within each season and identifies two different organizational patterns. This indicates that the dynamic changes of virtual vitality are reflected not only in differences in intensity at different time points but also in clear processes of formation, enhancement, decline, and maintenance.

A further comparison across seasons shows that virtual vitality is overall more stable in winter, while summer shows relatively greater variation. This is consistent with the finding of existing studies that “higher urban vitality is often accompanied by stronger persistence or stability” [62], indicating that it is also valid in the dimension of virtual vitality, thereby extending the related discussion from general urban vitality to the digital representation level.

The semantic analysis results show that online discussion content not only reflects users’ different perceptions of spatial scenes but also affects the sustained maintenance of virtual vitality. Existing studies have mainly used social media texts to identify city image, visual preference, and emotional characteristics [63,64], while this study further finds that the spatial orientation of online content is also an important factor affecting the stability of virtual vitality.

5.2. Spatial Organization of High-Vitality Stable Zones and Related Dynamic Zones

Hotspot analysis shows that there are clear differences in the spatial distribution of virtual vitality hotspots between summer and winter: winter hotspots are more concentrated around commercial centers and typical tourism nodes, while summer hotspots are more distributed around outdoor activity places with stronger openness. This result overall continues the understanding of existing studies on the seasonal differences of virtual vitality [65]. However, existing studies mostly remain at the comparison of vitality states at different time points or in different seasons [34]. Although they can reveal differences in the pattern of virtual vitality, they can hardly further identify which spaces can continuously maintain high vitality during seasonal transition, and which spaces show growth, decline, or persistently low vitality. On this basis, this study further uses the bivariate local spatial autocorrelation method and combines it with hotspot analysis results to identify four types of dynamic zones, HH, LH, HL, and LL, thus advancing the study of virtual vitality from the comparison of static patterns to the dynamic spatial identification of sustained maintenance states. Specifically, this study further defines high-vitality stable zones, vitality-enhancing zones, vitality-declining zones, and low-vitality stable zones in the seasonal transition process of virtual vitality, thereby providing an identification framework for understanding the sustained maintenance of online attention and its spatial supporting basis. The results show that HH zones are concentrated in the center of the study area and present a clustered agglomeration pattern, representing stable spaces where virtual vitality remains at a high level in both summer and winter. LH zones are mostly distributed around HH zones, reflecting potential growth spaces that are shifting from low to high levels and are easily driven by the diffusion of core zones. HL zones show a relatively scattered and irregular pattern, corresponding to spatial units that are stronger in summer and weaker in winter, indicating that their vitality depends more on outdoor activities and seasonal conditions. LL zones are mainly concentrated in the outer edge of the study area, indicating that there is a clear center–periphery differentiation of virtual vitality within the city. Overall, the seasonal transition of virtual vitality is not a simple replacement of hotspot locations, but is reflected in the different organizational positions and functional roles of different types of spatial units in the interaction between virtual and physical space, thereby extending the static description of existing studies mainly based on hot spot distribution to the dynamic spatial depiction of the sustained maintenance and transition paths of virtual vitality.

5.3. Formation Mechanism of High-Vitality Stable Zones

Existing studies have mostly explained the formation mechanism of urban vitality or virtual vitality from aspects such as the built environment, functional mix, and facility supply [30,31], but lack exploration of the formation mechanism of high-vitality stable zones of virtual vitality. Based on the identification results of HH zones, this study further explores the influencing factors in the formation of stable high-virtual-vitality areas. In the setting of influencing factors, existing studies have mostly described physical vitality space by overall indicators such as POI density or functional mix [48], making it difficult to reflect the differentiated effects of different facility types and spatial structural characteristics. This study further refines physical vitality spaces into different facility types and spatial structural characteristics for analysis.

Specifically, in terms of functional composition dissimilarity, the compositional dissimilarity of cultural and shopping functions shows a significant positive effect on high-vitality stable zones. This result indicates that the concentration or scarcity of a certain type of function within a local area can enhance the recognizability of the scene, affect individual spatial choice and activity behavior, thereby promoting the sustained formation of virtual vitality. When a certain type of function presents clustering in a local area, it often means that the area has formed an activity core dominated by this function, enhancing the overall recognizability and thematic perception of space through the continuous distribution of functional elements and consistent scene expression, making individuals more likely to stay continuously and engage in repeated consumption, thereby increasing the frequency of use and activity intensity of the space; when a certain type of function is relatively scarce, it is more likely to attract individuals to specifically go there due to its uniqueness or necessity attributes, triggering purposeful visit and experience behavior. Although these two differentiated manifestations follow different paths, both enhance the recognizability of the area and affect individual spatial choice and stay, thereby making it more likely to become source spaces for continuous content production and dissemination.

At the same time, the high importance of distance to activity center (tourism) indicates that spatial units closer to tourism activity centers are more conducive to the formation of high-vitality stable zones. Existing studies have pointed out that tourism spaces usually have strong communication potential and narrative attributes, and are more likely to become core carriers for the generation and diffusion of online content [66,67]. This study further finds that the online attraction formed by tourism space is not limited to the nodes themselves, but can spread to surrounding units through spatial proximity, and further affect the formation of a stable high-vitality state. Overall, differences in functional composition and proximity to tourism centers are not independent of each other, but jointly strengthen the identifiability of spatial units in the regional network, making them more likely to be continuously discovered, recognized, and spread in the process of virtual–physical interaction, thus providing a structural basis for the formation of high-vitality stable zones.

5.4. Planning Implications, Limitations, and Future Research

The results of this study provide a reference for the optimization of urban spatial structure and the sustained carrying and stable transformation of online attention in physical space, and further propose corresponding planning suggestions.

First, in terms of functional organization, it is necessary to enhance the recognizability of space by constructing functional combinations with clear themes, for example, by highlighting dominant functions such as culture or commerce within a local area, so that the space forms clear functional characteristics in the region, thereby making it easier to be perceived, recognized, and continuously attract activities.

Second, in terms of the configuration of daily service facilities, the promoting effect of basic facilities such as commerce and catering has thresholds, and a greater quantity is not necessarily more conducive to the stable high-vitality state of virtual vitality. Taking the study area of this paper and the 500 m × 500 m grid scale as an example, when catering facilities reach about 10–12 and shopping facilities reach more than about 20, a relatively sufficient promoting effect can be formed. This result can serve as a reference threshold for refined configuration in similar high-density urban areas, but its applicability still needs to be further verified under different urban types and spatial scales.

Finally, in terms of tourism spatial organization, it is necessary to construct a structure of “proximity strengthening–peripheral support” around core nodes. As areas closer to tourism centers are more conducive to the aggregation of online attention, planning should strengthen the continuity and accessibility around the core and guide the flow of people to diffuse toward peripheral areas through pedestrian paths and public nodes. At the same time, social media platforms can be used as auxiliary evaluation tools to continuously monitor the spatial distribution and dynamic change of online attention, and identify spatial units with development potential and emerging vitality nodes, thereby providing a basis for functional guidance and planning intervention of related spaces, and achieving coordinated optimization of physical space and digital space.

Despite these important findings, this study still has some limitations. To ensure data reliability, the check-in data collected in this study are limited and are confined to Harbin. While this helps improve the relevance of the data, it also restricts the sample size and cannot represent the characteristics of virtual vitality in different urban contexts. Second, the TikTok short-video data shows greater response to visually distinctive spaces. Therefore, there is still a certain difference between these data and the intensity of actual offline activities [68,69]. At the same time, TikTok user groups are mainly composed of younger populations [70], and the activity information of middle-aged and elderly groups as well as some non-platform users is difficult to be effectively captured, which means that the virtual vitality depicted in this study has, to a certain extent, the problem of uneven population coverage, thereby affecting the comprehensive representation of the overall urban vitality pattern.

Future research can further expand in two aspects: research context and data sources. In terms of research context, future studies can select multiple cities with relatively stable climates and more evenly distributed online attention throughout the year for comparative analysis, in order to test the applicability and differences of the relevant conclusions in different types of cities. In terms of data sources, this study is mainly based on data from TikTok, and the data coverage is relatively limited. Future studies can combine different social media platforms (such as Weibo, Xiaohongshu, and Facebook) to collect spatial behavior information of groups with different ages, genders, and cultural backgrounds, and further verify the robustness of the relevant mechanisms.

6. Conclusions

This study focuses on the dynamic change of virtual vitality, spatial association analysis, and influence mechanism modeling, constructs a multi-level analytical framework, and systematically reveals the spatial characteristics of high-vitality stable zones of virtual vitality and their relationship with the spatial structure of physical vitality spaces. The main conclusions are as follows:

(1)

Theoretical implications

This study shifts the research focus from the general change process of virtual vitality to the specific phenomenon of the stable high-vitality state, revealing its change characteristics in the temporal dimension and its stable maintenance mechanism. The study shows that the change process of virtual vitality is not completely random, but to a certain extent presents identifiable change patterns and forms spatial patterns with structures in different time periods. Among them, high-vitality stable zones (HH zones) form the core, surrounded by vitality-enhancing zones (LH zones), while vitality-declining zones (HL zones) and low-vitality stable zones (LL zones) present scattered and marginal distributions, respectively.

(2)

Methodological contributions

This study proposes, in terms of method, an integrated analytical framework starting from the dynamic change of virtual vitality and combining spatial pattern identification and mechanism explanation. This method, on the basis of characterizing the temporal variation characteristics of virtual vitality, can identify stable spatial structures and reveal the nonlinear relationship between them and the characteristics of physical vitality spaces. The results show that this method can distinguish the importance of different spatial elements for the formation of high-vitality stable zones and their nonlinear response characteristics, and realizes the systematic depiction of “change–pattern–mechanism” of virtual vitality.

(3)

Practical implications:

The results of the Random Forest model show that the formation of high-vitality stable zones relies more on structural characteristics with strong spatial recognizability rather than the accumulation of a single function quantity. Among them, distance to activity center (tourism) and functional composition dissimilarity (culture, shopping) have significant effects on the formation of high-vitality stable zones. This means that in the process of urban spatial optimization, priority can be given to strengthening the spatial organization and accessibility of core activity nodes, optimizing the distance relationship between key functions and core nodes within a certain range, and at the same time enhancing the degree of differentiated combination of cultural and commercial functions to improve the recognizability and attractiveness of space, thereby guiding the continuous transformation of online attention into physical space.

Overall, on the basis of revealing the dynamic change process of urban virtual vitality, this study further identifies stable high-vitality areas and explains their formation mechanism from the integrity of the physical vitality spatial structure. The results provide a new spatial perspective for understanding the process of the continuous accumulation of online attention in urban space and also offer practical planning suggestions for the transformation of online attention into physical space. It should be noted that this study is based on a single city and TikTok data, and its generalizability requires further validation with broader contexts and multi-source data.