Tourist flow, within the scope of “flow research” (including goods flow, traffic flow, tourist flow, information flow, etc.), was closely related to “mobility” (Zhong et al., 2010). Because tourism resources are generally immobile and unevenly distributed in space, tourists generate “flow” between different attractions to meet their own recreational needs. This phenomenon of tourists’ migration among scenic spots was called tourist flow (Zhang et al., 2005). The study of tourist flow began as early as the 1960s, mainly focusing on the spatial pattern and impact of tourist flow (Williams and Zelinsky, 1970). China’s research on this issue emerged at the end of the last century, initially focusing on the system, spatial distribution (Ma and Li, 2000; Zhang, 2000), and spatial structure (Wu and Bao, 2005; Zhong et al., 2009) of tourist flow. The construction and definition of tourism flow system were also still hot topics of research (Yuan, 2005; Xue, 2006) at that time. Later, the research focused on the spatial effect (Ma et al., 2008; Wang et al., 2010), spatial distribution characteristics, spatial transfer (Liu and Ma, 2008; Zheng et al., 2010), and flow quantity and quality analysis of tourist flow (Wang and Wu, 2016; Kang et al., 2020), mainly concentrating on the characteristics of spatiotemporal distribution (Yan et al., 2017) and network structure of tourist flow (Wang et al., 2023). Tourist flow had been widely valued by the academic community (Wang et al., 2022a). To this day, tourist flows have been widely valued by the academic community, with continuous in-depth research making certain progress in aspects such as spatial patterns, impacts, system construction, spatiotemporal distribution characteristics, and network structures. However, there is a lack of in-depth analysis of the dynamic changes in tourist flows and individual behaviors, and discussions on the social motives and psychological mechanisms behind tourist flows are not deep enough. This has limited the precision of tourist flow forecasting and the formulation of management strategies.

Tourist flow not only referred to the spatial movement trajectories of tourists but also encompassed the collection of various economic and social impacts resulting from this movement. It also reflected the connections between tourism destinations. Various spatial scales had been involved in the research on the spatial structure of tourist flow networks. The research questions included the structural characteristics (Sun et al., 2023), influencing factors (Xu et al., 2018; Li et al., 2021), evolutionary mechanisms (Yang and Wu, 2015), spatial patterns (Wang et al., 2020), comparative analysis of urban tourist flow network evolution (Yang and Wu, 2015), and cross-over studies with other societal hotspots (Cheng and Jia, 2020), etc. Various research methods were employed, such as GIS spatial analysis (Wang et al., 2021; Sun et al., 2023), social network analysis (Fang et al., 2023), QAP method (Zhou et al., 2020), etc. In terms of data acquisition, early studies heavily relied on survey questionnaires and official statistics, facing challenges in terms of quantity, adequacy, and timeliness. With the development of smart tourism, the mining of digital footprint data had made up for the defects of traditional data, and the ways of obtaining digital information such as online travel notes of tourism websites (Gholamhosseinzadeh et al., 2021), Baidu Index (Liang and Ma, 2023), Tencent population migration big data (Wu et al., 2020), Flickr website (Laura and Balzan, 2023), etc. have become new hotspots of attention. Overall, research on tourist flow networks covers aspects such as spatial structure, characteristics, influencing factors, evolutionary mechanisms, and urban evolution features. It is in a transitional phase from traditional data acquisition to the mining of digital footprint data, with the analysis of the complexity and dynamics of tourist flow networks not being comprehensive enough.

Change point detection (CPD), as an emerging interdisciplinary research method, was often applied to the research fields of statistics, signal processing, time series, etc., and was most commonly used for the prediction and anomaly detection of time series, which specifically referred to finding the smallest set of breakpoints in the given time series data, where the distribution or statistical properties of the data underwent a significant change at each breakpoint.

Assuming there was an ordered data sequence: y₁, ${{y}_{2}},\cdots,{{y}_{n}}$, a change occured at some point in time, resulting in different statistical properties before and after the change, this point was considered a change point (Wang and Huang, 2023). Several advanced algorithms had been proposed, mainly divided into three categories: statistical methods represented by autoregressive models (AR), moving average models (MA), pruned exact linear time (PELT), and simple exponential smoothing (SES); classical machine learning methods such as K-means clustering - subsequence time series clustering (STSC) (Keogh and Kasetty, 2003), one-class support vector machine (OC-SVM) (Tax and Duin, 1999), and local outlier factor (LOF) (Breunig et al., 2000); and deep learning methods, mainly including multilayer perceptron (MLP), residual neural network (Resnet) (He et al., 2016), and long short-term memory (LSTM) networks (Van Houdt et al., 2020), which widely were applied in various fields. With the maturity of the models and methods of complex network analysis (CNA), scholars had conducted extensive research in areas such as global transportation network characteristics (Del Mondo et al., 2021; Peng et al., 2021), landscape ecological security (Zhou et al., 2021), evolution of commodity trade patterns (Jiang and Wang, 2020), and more. The research on tourist flow network at the international (Wang et al., 2022b), domestic (Zhou and Xu, 2019) and urban agglomeration levels (Wang and Liu, 2022; Zhang and Yuan, 2023), a series of achievements had been made (Zhou et al., 2023). However, there is no research on applying the CPD to the tourism market stage division. Through this method, it is expected to identify the key time points more accurately in the evolution process of the tourist flow network and reveal the evolution process of the tourism network structure and the overall network characteristics. However, there has not yet been any research applying change point detection methods to the segmentation of tourism market stages. Through this method, it is expected to more accurately identify the key time points in the evolution of tourist flow networks, thereby revealing the evolution process of the tourist network structure and the overall network characteristics.

Currently, most of the research on tourist flow network relies on web text mining of tourism information data. Research methods included social network analysis (SNA) (Ren et al., 2023), Geography information system spatial analysis (GIS), Girvan-Newman algorithm (GN) (Zhang et al., 2023), among which social network analysis method has the largest proportion.

SNA, as a classic method of social relation research, mainly focused on the relationship and social structure (Zeng and Yu, 2022) between individuals and emphasizes revealing the group relationship and individual position in these relationships in the society. In SNA, more attention was paid to the relationship between members, and the individual attributes of network members were easily ignored. However, complex network analysis (CNA) can more widely study the general properties of network structure, including physical system, biological system, and social system, etc., and included various connection relationships between nodes, not limited to social relations. Applying CNA into the research of tourist flow network can use more types of measurement indicators, including degree distribution, clustering coefficient, network diameter, etc., to describe the structure and behavior of the network more comprehensively.

In summary, although there were abundant research achievements on tourist flow network, few studies had applied complex network methods to urban scale, and previous studies mostly divided time nodes by events, paying little attention to the natural evolution of the network. Tourist flow network is a dynamic system, using CPD to divide the time series can improve the accuracy of the time series division, and help to identify the turning points in the network time series, thus more accurately depict the spatial-temporal evolution characteristics of the network. During the outbreak of the COVID-19 pandemic, the tourism industry was severely impacted, and the radius of people’s travel was greatly restricted. Currently, there has been no scholarly research comparing the changes in the tourism flow network before and after the pandemic. With the pandemic factor superimposed, the changes in Shanghai’s tourism flow network from both spatial and temporal perspectives are of practical significance for further understanding the resilience of Shanghai’s tourism industry development and for more scientifically planning and distributing tourism resources in the post-pandemic era. Therefore, this paper collects the footprint data of online travelogues in Shanghai over the past decade, applies change point detection and complex network analysis methods, constructs a spatiotemporal evolution model of Shanghai’s tourism flow network, reveals the structural characteristics and spatiotemporal evolution laws, portrays its long-term trends and short-term fluctuation characteristics in evolution, and explores the endogenous driving forces of its evolution, providing more scientific and effective support for the spatiotemporal planning and management of tourism in Shanghai.

Shanghai, as the economic, financial, and trade center of China, is also a renowned tourist city in the country. It possesses unique advantages in terms of location, transportation, resource endowment, factor markets, and consumer markets for the development of the tourism industry.

Shanghai not only boasts rich cultural and tourism resources, including Red Culture, Jiangnan Culture, and Shanghainese Culture, but also gathers high-quality resources in finance, technology, and talent, along with numerous outstanding tourism enterprises. There are currently 130 national A-level tourist attractions, 11 national historical and cultural towns, 3449 immovable cultural relics, 1058 excellent historical buildings, and 44 historical preservation areas. Additionally, the city is home to 153 museums, 94 art galleries, 217 theaters and new performance spaces, and 385 cinemas.

In 2019, Shanghai received 361.4051 million domestic tourists throughout the year, with a total revenue of 552.2 billion yuan, with the tourism sector contributing over 6% to the city’s GDP. Despite the impact of the COVID-19 pandemic, Shanghai welcomed 236 million domestic tourists in 2020, generating 280.95 billion yuan in domestic tourism revenue. These figures represent recovery rates of 65.3% and 58.7% compared to 2019 levels for tourist numbers and revenue, respectively. Notably, Shanghai’s recovery rates exceeded the national average. In the first half of 2023, Shanghai hosted 144 million domestic tourists, with revenue totaling 172.4 billion yuan. These numbers indicate growth of 55% and 69% compared to 2020, respectively. Shanghai’s tourism industry had rebounded to 83% and 78% of the levels ^①

① Data source: https://whlyj.sh.gov.cn/tjzl/20210409/7e13241f0c8c4e7aad55bc0387c678f1.html.

during the same period in 2019, respectively, surpassing the national average recovery rates by 6% and 8%, respectively. These statistics highlighted Shanghai’s resilience and efforts in revitalizing its tourism sector despite the challenges posed by the pandemic. These advantages demonstrate that Shanghai, as a vibrant and modern tourist destination, maintains a unique appeal to visitors despite the challenges faced by the tourism market due to external factors.

Shanghai’s tourism industry leads the nation and serves as a model for other domestic cities (Li and Qu, 2021). Therefore, this study selects Shanghai as the research subject, intending to reveal the structural characteristics and evolution patterns of the STFN. The findings will contribute to the theoretical basis for optimizing the spatial layout of urban tourism.

The data collection involved mining online travelogue data from both Xiaohongshu (Little Red Book) and Qunar, two prominent platforms. Xiaohongshu, a rising social e-commerce platform in China, primarily attracts a young user base, boasting a current monthly active user (MAU) count of 260 million. In contrast, Qunar is a well-established domestic online travel service platform with a large user base. Filtering travelogues with Shanghai as the destination on the two platforms from 2014 to 2023, a total of 4326 footprint records after cleaning had been obtained. The basic geographical data was sourced from the National Basic Geographic Information Center, while the geographical coordinates of tourism nodes were obtained from the Gaode Maps API, using the WGS1984 coordinate system. Annual data on the number of domestic tourists received by Shanghai was derived from the “Shanghai Statistical Yearbook” and the Shanghai Tourism Industry Statistical Bulletin, covering a period from 2014 to 2022.

The STFN is constructed based on visitor footprints. This network is represented as an undirected weighted graph denoted by G=(V, E, W), where $V=\{{{v}_{1}},{{v}_{2}},\cdots,{{v}_{n}}\}$ represents the set of nodes, $E=\left\{ {{e}_{ij}} \right\}$ represents the set of edges, and $W=\left\{ {{w}_{ij}} \right\}$ represents the set of weights. Here, ${{e}_{ij}}$ indicates whether there is a connection between node ${{v}_{i}}$ and node ${{v}_{j}}$; if a connection exists,${{e}_{ij}}$=1; otherwise, ${{e}_{ij}}$=0. $W=\left\{ {{w}_{ij}} \right\}$ represents the set of weights, quantified by the number of likes received in travelogues. This network structure aids in identifying key nodes in tourist activities, understanding the relationships and connection strengths between nodes, and exploring the evolution of the tourist flow network over time.

Existing evaluations of univariate time series data indicates that statistical methods perform best in detecting point anomalies and collective anomalies. Not only do these methods provide more accurate anomaly detection, but they also execute faster, with shorter training and prediction times (Braei and Wagner, 2020). One notable statistical algorithm is the Pruned Exact Linear Time (PELT) algorithm proposed by Killick et al. (2012). This dynamic programming-based statistical method identifies structural change points in non-stationary sequence data. It achieves this by minimizing a cost function to determine the positions of change points. The expression is as follows:

where, Cost(t) represents the total cost of a split at position t, Cost(s) represents the total cost of a split at position s, Penalty(s+1, t) is the penalty term between positions s+1 and t, which is typically associated with the magnitude of the change between segmentation points.

Since the 1990s, complex network theory has rapidly developed and permeated various disciplines. Many scholars have applied it to model the tourist flow network, aiming to reveal the underlying mechanisms within systems. These models provide insights for data analysis, system optimization, and decision-making (Table 1). By building tourist flow networks connecting numerous tourist nodes, complex real- world phenomena can be simulated and explained (Zhang and Liu, 2014; Zhang and Liu, 2015).

Figure 1 illustrated the phase division of the Shanghai tourist market. It can be observed that the previous phase remained stable and orderly, while the subsequent phase experienced brief fluctuations followed by gradual recovery. The number of domestic tourists grew steadily during the previous phase, indicating that the scale of the tourist network continued to expand, with overall growth being relatively orderly and stable. However, upon entering the subsequent phase, the network’s complexity and uncertainty became evident. Disruptions caused by the COVID-19 pandemic led to a sharp decline in tourist numbers, increasing the overall instability of the tourism system. In 2021, as tourist numbers began to rise, the Shanghai tourist market faced a recovery, and a new evolutionary trend was budding.

The cumulative probability distribution of the two phases of the STFN was calculated (as shown in Figure 2) and compared it with an equivalent-sized (same number of nodes and edges) random network. The cumulative degree distribution of the STFN for both periods, when plotted on a log-log scale with logarithmic binning, approximately followed the scale-free characteristics typical of complex networks. In contrast, the random network exhibited a power-law distribution, highlighting significant differences between the two.

Quantitative indicators related to the small world characteristics of the STFN were statistically analyzed (as summarized in Table 2) (Barrat and Weigt, 2000). The values of $\sigma $ and $\omega $ were obtained by averaging results from ten random generations, and their standard deviation and coefficient of variation were calculated (Telesford et al., 2011). According to the table,$\sigma $>1 and $\omega $ falled within the range of (-1, 1), both standard deviations were below 10%, and the coefficients of variation were below 15%, confirming the validity of the computed results. In summary, both networks exhibited small-world network characteristics, indicating that the STFN possessed small-world characteristics and approximated a scale-free network, aligning with the fundamental features of complex networks.

Key nodes were typically nodes within a network that hold significance in maintaining stability and connectivity. These nodes played a crucial role in the overall structure and functionality of the network. By identifying and analyzing the characteristics of these nodes, a better understanding of the network’s structure and functions could be achieved (Loureiro et al., 2020). The weighted degree, closeness centrality, intermediate centrality, and eigenvector centrality of the top 10 tourist destinations in the STFN were calculated (as shown in Table 4) to analyze the key nodes changes of the network.

From the Table 4, it can be observed that:

(1) Various nodes experienced rapid growth in their weighted degrees, and Anfu road, emerging as a popular node in recent years, had standed out. This indicated that while the area around the Bund and Nanjing road walkway remained the core of the Shanghai tourism network, Anfu road’s influence had significantly increased due to evolving consumer demands.

(2) The centrality indicators for most nodes remained relatively stable, suggesting the network structure and the function of the key nodes were relatively stable, and traditional tourist destinations maintained consistent popularity levels.

(3) The Bund-Nanjing road walkway consistently served as a vital network resource hub. However, the strengthening intermediary effects of nearby ancient towns and their elevated status as “critical channels” had shifted the landscape. Nanxiang ancient town and Qibao ancient town, representing innovative models of cultural and tourism integration, played pivotal roles in enhancing network connectivity.

(4) In addition to traditional tourist destinations, the centrality of nodes such as the Shanghai world expo park and the Shanghai museum had further increased. This suggested that cultural nodes have become new popular choices, and the demand for cultural experiences had gradually become one of the essential factors influencing destination choices for tourists.

To further explore the spatial evolution characteristics of the STFN, ArcGIS 10.8 was used to transform the topological relationships of the Shanghai tourism network into spatial connections. The strength of the Shanghai tourist network linkages during the period 2014-2023 were sorted and visually represented in space, as shown in Figure 3. Referring to relevant threshold research results (Wu et al., 2015) and considering the quantity of connections in the STFN, linkages were extracted for 0-0.25%, 0.25%-0.5%, 0.5%- 1.0%, 1.0%-2.5%, 2.5%-5.0%, and 5.0%-10% of the main contact strengths. These were then merged and simplified into three types of connection networks: core network, skeleton network, and regional network (Jiang et al., 2023).

The connections within the core network had significantly increased, indicating connections between core nodes gradually strengthened, leading to the emergence of new popular tourism nodes. Initially, network connections were concentrated in the central area around the Bund. Over time, the network expanded southeastward, ultimately forming a pattern with the Bund as the core and Shanghai Disney resort as a secondary center.

The weighted degrees of various skeleton network nodes had increased to varying extents. This reflected the rising importance and influence of these nodes within the tourist flow network. The overall structure of the network remained relatively stable, exhibiting an inverted “V” shape. The southeastward expansion trend began to emerge.

The tendency for expansion was more pronounced within the regional network, with network linkages gradually extending from the central core nodes towards the southwest and southeast, forming an outward-radiating network. Simultaneously, this expansion extended beyond core-to- core connections and included links between core nodes and adjacent non-core nodes, such as the relationship between the world expo park and Century park. Nodes like the Shanghai astronomical museum and Dishui lake had formed relatively independent pairs of connections, which was due to their geographically distant locations from the city center.