Inter-Provincial Similarities and Differences in Image Perception of High-Quality Tourism Destinations in China

Abstract

With the rapid development of China’s tourism industry, the homogenization of regional tourism images has become a growing concern. To address this, this study quantifies the similarities and differences in tourism image perception across China’s 31 provinces, focusing on 350 5A-level destinations, analyzing 757,046 tourist reviews collected from Ctrip.com in 2024. Using a three-dimensional framework (cognitive, affective, and overall image), we analyze social media data through natural language processing, random forest regression, and social network analysis. Key findings include the following: (1) most comments are positive, with Jiangsu and Chongqing showing high cognitive image similarity but low overall similarity; (2) cognitive image significantly impacts affective image, especially through unique tourism resources; (3) an inter-provincial similarity–difference matrix reveals significant perceptual differences among provinces. This study provides a novel methodological approach for multidimensional image evaluation and offers crucial empirical insights for regional policy-making aimed at optimizing land and tourism resource allocation, balancing regional disparities, and promoting sustainable land use and development across China.

1. Introduction

In recent years, China’s tourism industry has experienced rapid growth, establishing itself as one of the largest tourism markets globally. This trend mirrors the development in many countries where tourism plays a significant role in national economies, such as Cyprus and Malta, highlighting the universal importance of sustainable destination management. With continuous socioeconomic progress and rising living standards, tourism—often referred to as a “happiness industry”—has become an integral part of daily life. To ensure the sustainable development of the sector, the Chinese government has implemented a series of policies promoting high-quality tourism development. Notably, the 14th Five-Year Plan for Tourism Development explicitly calls for the transformation of the tourism industry towards a more sustainable and high-quality model, emphasizing optimized resource allocation, service enhancement, and spatial planning. This strategic direction aligns with global efforts to balance tourism growth with environmental and cultural preservation, providing a critical context for upgrading China’s tourism sector.

In China, tourism destinations are officially classified into five quality levels, from A to AAAAA (5A), based on rigorous criteria such as service quality and visitor appeal (mct.gov.cn). For the purpose of this study, we operationally define ‘high-quality tourism destinations’ as these national 5A-level sites. This operationalization is well justified for two primary reasons. First, these sites represent the pinnacle of tourism excellence in the country and serve as key drivers of regional tourism economies [1,2]. Second, due to their significant scale, comprehensive facilities, and immense influence, they often function as standalone destinations in their own right, attracting large volumes of both national and international tourists. Accordingly, to ensure a comprehensive and nationally representative analysis, our study includes 350 officially designated 5A-level destinations. However, the increasing number of 5A-level tourism destinations has led to concerns over homogenization, with many destinations exhibiting similar characteristics in terms of tourism resources and services. This not only diminishes the uniqueness of individual destinations but also reduces the overall diversity of the tourism experience [3].

Understanding the interplay between similarity and difference in tourism image perception is critical for sustainable regional development. On one hand, excessive similarity can lead to market saturation, limit tourist choices, and result in inefficient resource allocation [4]. On the other hand, stark differentiation, while enhancing individual competitiveness, may hinder regional branding and collaboration. Therefore, striking a strategic balance is essential for both coordinating regional tourism efforts and catering to diverse tourist preferences [5]. However, current tourism management is predominantly confined within provincial boundaries, creating a major obstacle to such coordination. Conducting large-scale research on cross-provincial tourism destinations is therefore essential for understanding inter-provincial relationships and promoting regional coordination [6]. By systematically quantifying similarities and differences in image perception, policymakers and tourism stakeholders can facilitate greater integration of tourism industries, optimize resource allocation, and foster balanced development [7]. This approach not only strengthens tourism branding and enhances regional image positioning but also removes spatial barriers, expands tourists’ choices, and promotes new forms of tourism, such as cross-provincial self-driving tours [8].

Image perception research is predominantly grounded in the “cognitive–affective” theory from psychology [9]. Baloglu first applied this framework to tourism destination image studies, proposing that image perception is a multidimensional construct comprising cognitive, affective, and overall dimensions. The cognitive image pertains to tourists’ functional and knowledge-based perceptions, while the affective image reflects their emotional responses to a destination. The overall image is a holistic synthesis of these two dimensions [10]. Particularly, the cognitive image exerts a significant influence on the affective image, because emotions are a key driver of tourist satisfaction, loyalty, and decision-making [11]. By examining the contribution of cognition to emotional perceptions, researchers and practitioners can gain valuable insights into the psychological mechanisms that underpin tourists’ attitudes and behaviors [12]. However, most existing studies have analyzed image perceptions from either a single or dual-dimensional perspective, lacking an integrated framework that fully explores the interrelationships between cognitive, affective, and overall image perceptions. Therefore, building on the dimensions of cognitive, affective, and overall perceptions, the introduction of research on the relationship between cognition and affective holds significant theoretical importance. It not only advances the study of tourism image perception but also expands the scope of the “cognitive–affective” theory.

The proliferation of user-generated content on online review platforms offers an unprecedentedly rich data source for understanding genuine tourist perceptions [13]. This paradigm shift in data availability calls for advanced computational methods. This study, therefore, leverages an integrated framework combining natural language processing (NLP) to analyze textual data, random forest regression to model the complex relationships between image dimensions, and social network analysis (SNA) to map inter-provincial similarities and differences [14,15,16]. This computational approach enables a more objective and nuanced analysis of tourism image on a large scale than has been possible with traditional methods.

Therefore, this study aims to address the aforementioned research gaps. The specific objectives are threefold: (1) investigate the multidimensional structure of tourism image perceptions (cognitive, affective, and overall) across China’s provinces; (2) explore the cognitive–affective relationship in tourism image perception, revealing how cognitive evaluations shape affective responses; (3) construct and analyze an inter-provincial similarity–difference matrix for high-quality tourism destinations. Theoretically, this study expands the tourism image perception theory by integrating the “cognitive–affective–overall image” three-dimensional framework through big data methods. Practically, this study establishes an inter-provincial tourism destination similarity–difference matrix, offering empirical evidence for regional tourism resource integration and high-quality development. It supports China’s tourism industry in advancing regional collaboration and achieving high-quality development.

2. Data and Methods

2.1. Research Framework

Employing Python 3.8-based analytical tools, the research processes online review data from 350 5A-level tourism destinations across 31 provincial administrative regions in China, excluding Hong Kong, Macau, and Taiwan (Figure 1). Utilizing SnowNLP, Latent Dirichlet Allocation (LDA), Jieba word segmentation, and keyword extraction, the investigation explores tourists’ perceptions of 5A-level tourism destination images within the cognitive–affective–overall framework. Then, random forest regression is employed to examine the relationship between cognitive and affective images. By integrating the aforementioned affective, cognitive, overall, and cognitive-to-affective image, a comprehensive similarity matrix is derived. This matrix emphasizes the static and local similarities among provinces. Meanwhile, the study utilizes the comprehensive similarity to derive the network relationships of image perception among provinces through SNA, employing edge weights to represent dynamic and global similarities. Finally, the inter-provincial similarity–difference matrix is generated by integrating the comprehensive similarity matrix and the network edge weight matrix.

2.2. Data

For this study, we collected tourist review data from Ctrip [17], one of China’s largest and most widely used online travel platforms. The data was programmatically crawled using Python, targeting all 350 5A-level tourism destinations across 31 provinces (excluding Hong Kong, Macao, and Taiwan). The platform’s extensive user base and large volume of user-generated content make it a highly valuable and representative data source for tourism perception studies [18]. It is important to note that these destinations are specific, officially designated scenic areas, not entire cities or regions. The detailed list is provided in Table S1, and their geographical distribution is shown in Figure 2. The list of 5A-level tourism destinations was obtained from the official directories published by provincial departments of culture and tourism (the statistical cutoff date is primarily concentrated in February 2024). It should be noted that there is not a one-to-one correspondence between 5A-level tourism destinations and those listed on the Ctrip platform. It is possible for one 5A-level destination to correspond to two Ctrip destinations and for several 5A-level destinations to correspond to one Ctrip destination. We have made every effort to collect the relevant data, but there are still nine destinations for which data is missing. Data collection was conducted from 28 November 2024 to 4 December 2024, covering a review period of ten years (16 October 2015 to 4 December 2024). In total, 757,046 original reviews were collected. To ensure data quality, the acquired reviews underwent a rigorous prepossessing procedure. This involved removing irrelevant columns; cleaning comment content by eliminating HTML tags, URL links, special symbols, and emoticons; as well as deleting empty or duplicate entries. After prepossessing, a total of 720,760 valid reviews remained for analysis. In this study, https://www.chiplot.online/(accessed on 4 December 2024) was utilized for chart creation.

2.3. Natural Language Processing

2.3.1. Sentiment Analysis

This study employs the SnowNLP library to assess the sentiment of the collected reviews. SnowNLP is an NLP tool based on Bayesian modeling, widely utilized for sentiment analysis in Chinese text, which aligns well with the objectives of this study [19]. For each review, the sentiment score is computed on a scale from 0 to 1, where 0 represents a highly negative sentiment and 1 represents a highly positive sentiment. Based on these scores, reviews are classified into three categories: positive (score > 0.75), neutral (0.25 ≤ score ≤ 0.75), and negative (score < 0.25). Additionally, the total count of reviews in each sentiment category is recorded, and an overall sentiment average is calculated.

2.3.2. Keyword Extraction

To extract keywords from the review data, this study utilizes Jieba for Chinese word segmentation and the Term Frequency–Inverse Document Frequency (TF-IDF) algorithm for keyword weighting [5,20]. The TF-IDF algorithm was specifically chosen because it effectively identifies words that are not only frequent within a review but also distinctive to it, which is crucial for capturing the unique characteristics of each destination’s image. For each review, the top five weighted keywords were extracted.

A keyword frequency table is then generated for each province by counting the occurrences of each keyword. The top 50 keywords for each province are manually categorized into six groups: tourist destinations, tourism facilities, tourism resources, tourists, affective vocabulary, and others. This classification aids researchers in focusing on the most relevant keywords. The detailed formulas for the TF-IDF calculation are available in the Supplementary Materials (Section S1).

2.3.3. Topic Modeling

To identify latent thematic structures in the reviews, this study applies the LDA topic modeling method [21]. LDA is suitable for research that requires automatic topic extraction and the discovery of hidden semantic structures from large volumes of text [22]. A crucial step is determining the optimal number of topics. To guide this, we trained models with 2 to 10 topics on the Beijing dataset, evaluating each with the topic coherence score (C_v). The analysis revealed two notable peaks: a significant initial peak at five topics (C_v = 0.4501) and the global maximum at eight topics. Although the eight-topic model scored slightly higher, it produced more granular and overlapping themes. For the purpose of a focused and streamlined provincial-level analysis, we selected the five-topic solution. This choice captures the most essential, distinct, and managerially relevant themes, representing the best balance between model coherence and thematic clarity for the scope of our study. This approach enables multi-processing, lexical prepossessing, word frequency matrix construction, topic extraction, and label generation. LDA facilitates the identification of overarching themes within the dataset while allowing for detailed keyword analysis at a micro-level, thereby refining the extracted thematic content.

2.4. Random Forest Regression

To analyze the relationship between keywords and sentiment scores, this study employs random forest regression, a machine learning algorithm capable of capturing complex, non-linear relationships through ensemble learning [23]. This model is particularly effective in handling high-dimensional and sparse textual data while assessing the influence of individual variables on sentiment scores [24]. Compared to traditional linear regression models, random forest provides superior predictive accuracy and adaptability to intricate data patterns.

The modeling process followed a standard machine learning workflow. First, the textual keyword data was pre-processed and transformed into a numerical feature matrix. The dataset was then partitioned into training and testing sets to build and validate the model. Subsequently, the random forest model was trained, and its hyperparameters were optimized to maximize performance. Finally, the model’s predictive accuracy was evaluated using Mean Squared Error (MSE) and R² scores, and the feature importance values were extracted to identify the most influential keywords affecting sentiment.

All technical implementation details, including the specific Python libraries and functions used, the vectorization parameters, the dataset split ratio, and the hyperparameter search space, are documented in the Supplementary Materials (Section S2).

2.5. Cosine Similarity

This study employs Cosine Similarity to measure the similarity of different variables across provinces. Cosine Similarity is a widely used metric that evaluates the angle between two vectors, effectively quantifying the degree of similarity between provincial feature sets [25]. The calculation is as follows:

Given two vectors A and B (which represent the feature vectors of two provinces), the Cosine Similarity is defined as

$$\mathrm{C}\mathrm{o}\mathrm{s}\mathrm{i}\mathrm{n}\mathrm{e}\mathrm{S}\mathrm{i}\mathrm{m}(\mathrm{A},\mathrm{B})={\displaystyle \frac{\mathrm{A}\ast \mathrm{B}}{‖\mathrm{A}‖‖\mathrm{B}‖}}$$

(1)

where

$\mathrm{A}\ast \mathrm{B}$ is the dot product of vectors A and B, calculated as

$$\mathrm{A}\ast \mathrm{B}=\sum _{\mathrm{i}=1}^{\mathrm{n}}{\mathrm{A}}_{\mathrm{i}}{\mathrm{B}}_{\mathrm{i}}$$

(2)

$‖\mathrm{A}‖$ and $‖\mathrm{B}‖$ are the Euclidean norms of the vectors A and B, respectively, calculated as

$$‖\mathrm{A}‖=\sqrt{\sum _{\mathrm{i}=1}^{\mathrm{n}}{\mathrm{A}}_{\mathrm{i}}^2}$$

(3)

$$‖\mathrm{B}‖=\sqrt{\sum _{\mathrm{i}=1}^{\mathrm{n}}{\mathrm{B}}_{\mathrm{i}}^2}$$

(4)

A similarity value closer to 1 indicates higher similarity between two provinces, whereas values closer to 0 indicate greater differences.

2.6. A Network Analysis of Inter-Provincial Similarities and Differences in High-Quality Tourism Destinations

To systematically analyze inter-provincial similarities and differences in high-quality tourism destinations, this study focuses on image perception among provinces. Initially, a mean analysis is conducted on affective, cognitive, overall, and cognitive-to-affective image similarities to derive a comprehensive image similarity matrix for each province.

Subsequently, SNA is employed to construct a network based on image similarity [26]. In this network, nodes represent individual provinces, while edges represent the strength of the connection between two nodes. Edge weights not only convey networked expression of similarity but are also influenced by the overall structure of the network. For instance, if a province exhibits high similarity with multiple other provinces, its centrality within the network (such as degree centrality or betweenness centrality) may be elevated, thereby impacting the entire network’s architecture.

Finally, based on the comprehensive similarity matrix and the edge weight matrix derived from SNA, the inter-provincial similarity–difference matrix is formed. The natural breaks method is employed to categorize inter-provincial relationships into five distinct classes:

Category I: High similarity–high edge weight, high similarity–medium edge weight, medium similarity–high edge weight.

Category II: Medium similarity–medium edge weight.

Category III: Low similarity–medium edge weight, low similarity–high edge weight.

Category IV: Medium similarity–low edge weight, high similarity–low edge weight.

Category V: Low similarity–low edge weight.

The integration of similarity and edge weights offers the advantage of capturing both the static local similarities and the dynamic global similarities among provinces, providing a more comprehensive analytical perspective. This approach allows for the meticulous categorization of inter-provincial relationships, facilitating the formulation of precise strategies.

3. Results

3.1. Affective Image Analysis

The affective tone of comments for high-quality tourism destinations in most provinces is predominantly positive, with the proportion of positive comments generally exceeding 70% (Figure 3). Among these, Shanghai (85.06%), Hainan (84.96%), and Ningxia (84.79%) report the highest proportions of positive comments, while Guizhou has the lowest (68.16%). Provinces with higher proportions of negative comments include Guizhou (19.24%), Yunnan (15.73%), and Shanxi (15.69%), while Ningxia exhibits the lowest proportion of negative comments (6.92%) (

Table 1. Evaluation of the number, proportion, and average scores of three types of comments by province.

Province	Number of Positive Comments	Percentage of Positive Comments	Number of Neutral Comments	Percentage of Neutral Comments	Number of Negative Comments	Percentage of Negative Comments	Average Score
Shanghai	7036	85.06%	487	5.89%	749	9.05%	0.875
Yunnan	19,496	73.26%	2929	11.01%	4186	15.73%	0.779
Inner Mongolia	8153	74.25%	1369	12.47%	1458	13.28%	0.793
Beijing	20,594	84.57%	1737	7.13%	2020	8.30%	0.874
Jilin	10,357	79.14%	1423	10.87%	1307	9.99%	0.832
Sichuan	28,063	77.54%	3705	10.24%	4422	12.22%	0.817
Tianjin	4305	74.33%	834	14.40%	653	11.27%	0.800
Ningxia	9469	84.79%	925	8.28%	773	6.92%	0.878
Anhui	19,886	77.91%	2971	11.64%	2666	10.45%	0.825
Shandong	24,392	75.12%	3526	10.86%	4552	14.02%	0.794
Shanxi	16,281	72.55%	2639	11.76%	3522	15.69%	0.773
Guangdong	20,625	77.26%	3359	12.58%	2713	10.16%	0.820
Guangxi	13,870	74.42%	2013	10.80%	2754	14.78%	0.789
Xinjiang	18,461	75.93%	2991	12.30%	2861	11.77%	0.809
Jiangsu	54,578	80.53%	6843	10.10%	6356	9.38%	0.845
Jiangxi	17,213	73.57%	2679	11.45%	3506	14.98%	0.781
Hebei	12,287	74.16%	2118	12.78%	2164	13.06%	0.793
Henan	26,669	80.31%	3155	9.50%	3382	10.18%	0.840
Zhejiang	39,253	80.10%	4771	9.74%	4983	10.17%	0.839
Hainan	17,175	84.96%	1526	7.55%	1514	7.49%	0.877
Hubei	30,636	75.55%	5089	12.55%	4826	11.90%	0.805
Hunan	20,638	76.92%	3021	11.26%	3171	11.82%	0.814
Gansu	9757	77.09%	1293	10.22%	1607	12.70%	0.813
Fujian	15,210	76.49%	2297	11.55%	2377	11.95%	0.811
Xizang	8998	81.62%	1009	9.15%	1017	9.23%	0.851
Guizhou	12,628	68.16%	2334	12.60%	3564	19.24%	0.735
Liaoning	12,900	80.04%	1669	10.36%	1548	9.60%	0.839
Chongqing	14,350	71.97%	2847	14.28%	2741	13.75%	0.777
Shaanxi	28,325	77.79%	3537	9.71%	4550	12.50%	0.818
Qinghai	5029	81.06%	587	9.46%	588	9.48%	0.848
Heilongjiang	5469	71.07%	1159	15.06%	1067	13.87%	0.772

). In terms of average scores, Guizhou (0.735) falls within the neutral range, while all other provinces score positively. Ningxia has the highest average score (0.878), followed by Hainan (0.877), Shanghai (0.875), and Beijing (0.874).

3.2. Cognitive Image Analysis

Cognitive image similarity among provinces is generally low (Figure 4a). However, Jiangsu, Chongqing, Hubei, and Henan demonstrate relatively high cognitive image similarity. Shanghai, in contrast, shows no cognitive image similarity with other provinces, suggesting substantial differences in its tourism experiences compared to other regions. The top 50 keywords for each province were categorized into six groups: tourist destinations, tourism facilities, tourism resources, tourists, affective vocabulary, and others (Table S2). Among these, tourism resources are the most diverse, reflecting tourists’ varied preferences for destinations and resources across different provinces. Regarding tourism facilities, tourists show strong recognition of convenient transportation options such as “sightseeing bus,” “cable car,” “electric cart,” and “boardwalk,” as well as supporting amenities like “tour guide,” “commentary,” and “commentator.” In terms of tourists, provinces like Shanghai, Jilin, Ningxia, Shandong, Guangdong, Henan, and Liaoning emphasize the importance of child-friendly features in tourism destinations. From the affective vocabulary, it is evident that tourists across provinces generally hold positive evaluations of tourism. However, cognitive differences exist among provinces concerning tourism resources, facilities, and affective vocabulary, with smaller differences in the perception of tourism resources (Figure 5).

3.3. Overall Image Analysis

The overall image similarity across most provinces is relatively low (Figure 4b). Provinces like Shanghai, Anhui, Guangxi, Zhejiang, and Guizhou show no overall image similarity with other regions. However, higher similarity is observed between Liaoning and Heilongjiang, Hainan and Inner Mongolia, and Hainan and Liaoning, suggesting shared characteristics in tourism resources or cultural features. The tourism labels of each province exhibit high diversity in overall image (

Table 2. Theme labels by province.

Province	Theme 1 Label	Theme 2 Label	Theme 3 Label	Theme 4 Label	Theme 5 Label
Shanghai	Wild Animals	Landmarks/The Bund/Lujiazui	Shanghai	Education/Science Popularization/China	Night View/Hours
Yunnan	Cable Car	Eastern District	Hot Springs	Scenic Area/Nature
Inner Mongolia	Erdao Bridge/Most Beautiful/Places	Recommendation	Time	Scenery
Beijing	Heshen	China/Beijing	Scenic Area/Convenience/Beijing	Tickets/Places/Attractions	Ancient China/Hall of Prayer for Good Harvests
Jilin	Children	Scenery/Changchun	Commentary/Places	West Slope/World/Attractions
Sichuan	World Heritage	Jiuzhaigou	Scenery/Attractions/Pandas	Snow Mountain/Guangwu	Commentary
Tianjin	Clay Figurine Zhang	Clay Figurine Zhang	Scenery/Scenic Area/Jixian	Children/Hiking/Cost-Effectiveness
Ningxia	Scenery/Service	Tickets	Experience/Heritage Site	Filming/Scenes/Zhenbeibao
Anhui	Scenic Area/Hours/Descent	Bodhisattva/Tiantai	Ancient Villages/Huizhou/Nanhu	Recommendation/Places
Shandong	Cable Car/Ropeway/Hours	Recommendation/Attractions/Liugong Island	Convenience/Qingdao/Evening	Confucius Mansion	Tickets
Shanxi	Buddha Statues	Shanxi/Architecture	Places	Queuing/Mianshan	Hukou Waterfall
Guangdong	Places/West Lake	Children	Time	Xiqiao Mountain/Huizhou	Environment/Stone Towers/Scenery
Guangxi	Scenery/Cross-Border/Scenery	Bamboo Raft/Landscape/Dock	Lijiang River/Night Tour/Duxiu Peak	Culture/Experience/History	Scenery/Tickets
Xinjiang	Scenery/Grand Canyon/Keketuohai	Scenery	Tianshan/Bayanbulak	Clear/Lakes	Time/Shuttle Bus
Jiangsu	Tickets	Giant Buddha/Places	Linggu/Meiling Palace/Nanjing	Suzhou/Yuantouzhu	Jinshan/Confucius Temple
Jiangxi	Convenience	World/Lushan/Scenery	China/Jiangnan	Mountain Climbing/Sanqingshan/Attractions
Hebei	Cable Car/Scenery	Scenery	Underground Palace/Mausoleum	Architecture/Summer Resort/Ancient City
Henan	Shaolin	Grand Canyon/Taihang	Tickets	Children/Places/Attractions	Shows/Evening
Zhejiang	Tickets	Ningbo/Yandang Mountain/Hometown	Hangzhou	Water Town/Scenic Area/Jiangnan Water Town	Qiandao Lake
Hainan	Sanya	Scenery	Sanya	Small Cave/Places
Hubei	Convenience/Fun/Recommendation	Park/Chibi	Tickets	Scenery/Suitable/Places	People/Dam/Scenery
Hunan	Glass/Mountain/Boardwalk	Convenience	Like/Recommendation	Scenery/Chairman Mao/Places	Night View
Gansu	Convenience	Dunhuang/Tickets	Kongtong Mountain/China	Desert/First Pass Under Heaven/Fortress
Fujian	Gulangyu/Wuyi Mountain	Attractions/Mazu	Rafting	Recommendation/Suitable/Fun
Xizang	Devotion	Linzi/Tibet	Monasteries/Tibetan Buddhism	Lake	Time
Guizhou	Attractions	Guizhou/Characteristics	Mountain Climbing/Weather	Guizhou/Magnificent	Time/Tianxingqiao/Hours
Liaoning	Convenience/Animals	Places/Scenic Area	Scenic Area	Places
Chongqing	Places	Attractions/Tourists	Chongqing/Attractions/Wulong	Waterfall/Baidicheng	Art/Grottoes/Beishan
Shaanxi	Mountain/Cable Car	Convenience/Attractions	Park/Bicycle/Lishan	Experience/Famen Temple/Architecture	Recommendation
Qinghai	Saltwater Lakes/Beautiful	Commentary	Temple/Gelugpa	Driver/Time	Niuxin Mountain/Faith
Heilongjiang	Ice and Snow/Songhua River	Tickets	5A	Places	Scenic Area/Scenery/Magnificent

). For example, natural landscapes include Yunnan’s “hot springs,” Xinjiang’s “grand canyon/Keketuohai,” and Qinghai’s “saltwater lakes.” Historical culture examples include Beijing’s “Hall of Prayer for Good Harvests/ancient China” and Shandong’s “Confucius Mansion.” Religious and spiritual culture is represented by Xizang’s “monasteries/Tibetan Buddhism” and “devotion.” Modern cities and service facilities are exemplified by Shanghai’s “the Bund/Lujiazui” and “education/science popularization.” These labels reveal unique characteristics of tourism resources across provinces and highlight tourists’ diverse needs for natural, cultural, and service-related experiences, offering a strong foundation for targeted marketing and resource optimization.

3.4. Analysis of the Relationship Between Cognitive and Affective Image

The results of the random forest regression model reveal the influence of cognitive images on the affective image of tourism destinations (

Table 3. Top 10 keywords contributing the most to affective scores by province.

Shanghai	Queuing	Too Many People	Ctrip	160	Buy Tickets	Oriental Pearl Tower	Go Up	ID Card	Animals	Science and Technology Museum
Yunnan	Lijiang	Ancient Town	Scenic Area	Inn	Ancient Town	Scenery	Eggs	Tickets	Hot Springs	Electric Cart
Inner Mongolia	Package Tickets	Scenic Area	Queuing	Scenery	Grassland	Ctrip	Collect Tickets	Attractions	Not Bad	Tickets
Beijing	ID Card	877	Ctrip	Great Wall	Queuing	Tour Guide	Buy Tickets	Cable Car	Echo Wall	20
Jilin	Ctrip	Queuing	Buy Tickets	Tickets	Book Tickets	Scenic Area	Tianchi	Not Bad	Changchun Film Studio	Scenery
Sichuan	Scenic Area	Collect Tickets	ID Card	Queuing	Ctrip	Scenery	Attractions	Cable Car	Dujiangyan	Tickets
Tianjin	Jianbing Guozi	Scenic Area	Not Bad	Scenery	Ancient Culture	Tianjin	60	Panshan	Queuing	Ear Hole Fried Cake
Ningxia	Scenic Area	Ctrip	Package Tickets	Tickets	Journey to the West	Shapotou	Queuing	QR Code	Play	Desert
Anhui	Collect Tickets	Scenic Area	Tickets	Ctrip	Not Bad	Scenery	Hongcun	Attractions	Cable Car	Huangshan
Shandong	Scenic Area	Not Bad	Queuing	10	Scenery	Penglai Pavilion	Attractions	Rafting	Tickets	Buy Tickets
Shanxi	Pingyao	Hukou Waterfall	Scenic Area	Ancient Town	Waterfall	Collect Tickets	Not Bad	Ctrip	Attractions	Tickets
Guangdong	Not Bad	Scenery	Queuing	Tickets	Scenic Area	Fees	Collect Tickets	Attractions	Ctrip	Buy Tickets
Guangxi	Lijiang River	Guilin	Elephant Trunk Hill	Scenic Area	Collect Tickets	Ctrip	Scenery	Waterfall	Bamboo Raft	Duxiu Peak
Xinjiang	Scenic Area	Shuttle Bus	Scenery	Queuing	Grassland	Collect Tickets	Attractions	Ctrip	Tickets	Too Hot
Jiangsu	ID Card	Scenic Area	Ctrip	Attractions	Queuing	Ancient Town	Collect Tickets	Tickets	Not Bad	Scenery
Jiangxi	Tengwang Pavilion	Queuing	Collect Tickets	Scenic Area	Scenery	Cable Car	ID Card	Not Bad	Ctrip	Attractions
Hebei	Scenery	Scenic Area	Not Bad	Ctrip	10	Buy Tickets	Attractions	Tickets	First Pass Under Heaven	Collect Tickets
Henan	Scenic Area	Collect Tickets	Attractions	Ctrip	Queuing	Cable Car	Tickets	Not Bad	Scenery	Parking Lot
Zhejiang	Scenic Area	ID Card	Tickets	Ctrip	Attractions	Collect Tickets	Ancient Town	Not Bad	West Lake	10
Hainan	Scenic Area	Ctrip	Electric Cart	Attractions	Queuing	Collect Tickets	Small Cave	Book Tickets	Wuzhizhou Island	Tour Bus
Hubei	Scenic Area	Not Bad	Scenery	Attractions	Collect Tickets	Waterfall	Ctrip	ID Card	Tickets	Tourists
Hunan	Queuing	Collect Tickets	Scenic Area	Yueyang Tower	Not Bad	Orange Island	Scenery	Attractions	Ctrip	Tianmen Mountain
Gansu	Scenic Area	Crescent Moon Spring	Jiayuguan	Tickets	Attractions	ID Card	Enter	Danxia	Mingsha Mountain	Queuing
Fujian	Scenic Area	Collect Tickets	Bamboo Raft	Ctrip	Not Bad	Wuyi Mountain	Gulangyu	Attractions	Tickets	Tulou
Xizang	Lhasa	Reservation	Tickets	Potala Palace	Tour Guide	Jokhang Temple	Queuing	Commentary	Ctrip	Attractions
Guizhou	Waterfall	Scenic Area	Queuing	Tianxingqiao	Huangguoshu	Scenery	Not Bad	Attractions	Ancient Town	Collect Tickets
Liaoning	Ctrip	Queuing	Buy Tickets	Scenery	Scenic Area	Collect Tickets	Water Cave	Not Bad	ID Card	Book Tickets
Chongqing	Not Bad	Scenery	Scenic Area	ID Card	Collect Tickets	Tiankeng	Ctrip	Queuing	Scenery	Hours
Shaanxi	Hukou Waterfall	Scenic Area	Collect Tickets	Waterfall	Buy Tickets	Cable Car	Ctrip	Commentary	Queuing	Tour Guide
Qinghai	Kumbum Monastery	Qinghai Lake	Scenic Area	Buy Tickets	20	Collect Tickets	Ctrip	Enter	Parking Lot	Tibetan Buddhism
Heilongjiang	Waterfall	Scenery	100	Not Bad	Scenic Area	ID Card	Ctrip	Tickets	Attractions	Sun Island

). Specific iconic attractions in certain provinces significantly impact affective images, such as “Lijiang” in Yunnan and “Pingyao” in Shanxi. Additionally, general terms like “scenic area,” “scenery,” and “attractions” also contribute substantially to affective images. Furthermore, regional cultural and tourism characteristics influence affective cognition. For example, Tianjin’s “jianbing guozi” (a local snack) symbolizes the personalized impact of cognitive images on affective images. The frequent appearance of the term “ID card” in comments about Shanghai and Beijing warrants attention from scenic area managers. Conversely, negative terms such as “queuing” appear in comments from over 20 provinces, highlighting overcrowding issues in high-quality tourism destinations. Notably, Shanghai, besides “queuing,” also features terms like “too many people,” which strongly affect tourists’ affective perceptions. The relationship between cognitive and affective images is relatively low in similarity across provinces, with the most significant contrast observed between Shanxi and Shaanxi (Figure 4c).

3.5. Inter-Provincial Similarities and Differences in High-Quality Tourism Destinations

This section explores the similarities and differences among high-quality tourism destinations across provinces through the analysis of a comprehensive similarity matrix, network edge weight matrix, and inter-provincial similarity–difference matrix. First, the comprehensive similarity matrix illustrates all the four similarities in image perception among provinces (Figure 6a). The results show that the similarity between most provinces is relatively low (>0.2). Notably, the similarity between Jiangsu, Henan, Hubei, and Chongqing exceeds 0.5, indicating high consistency in image perception. Second, the analysis of network edge weights unveils the dynamic and global similarities among various provinces (Figure 6b). Most edge weights between provinces range from 0.2 to 0.6. Provinces such as Jilin and Beijing, Shaanxi and Jilin, Gansu and Jilin, Gansu and Shaanxi, and Gansu and Beijing exhibit values greater than 0.5. In contrast, provinces like Guizhou and Xinjiang, and Guizhou and Guangxi show lower edge weights.

Finally, by combining the comprehensive similarity matrix and the network edge weight matrix, an inter-provincial similarity–difference matrix is formed (Figure 6c). The analysis indicates that high-quality tourism destinations’ image perceptions in China exhibit significant differences, with 276 pairs in Category V, 84 pairs in Category IV, 84 pairs in Category III, and 20 pairs in Category II. Only Beijing and Shaanxi fall into Category I. Category I includes provinces with high or middle comprehensive similarity and strong network connections (high or middle edge weights). Category II consists of provinces with middle comprehensive similarity and intermediate network connections (middle edge weights). Category III encompasses provinces with low comprehensive similarity but moderate-to-strong network connections (middle-to-high edge weights). Category IV involves provinces with high or middle comprehensive similarity but weak network connections (low edge weights). Finally, Category V comprises provinces with low comprehensive similarity and weak network connections (low edge weights). This classification provides a systematic understanding of inter-provincial image perception of high-quality tourism destinations dynamics, supporting targeted strategies for regional development and collaboration.

4. Discussion

4.1. Interpreting the Patterns of High-Quality Tourism Destination Images: Homogeneity in Diversity

Our analysis reveals a complex geography of tourism image perception in China, characterized by both distinctiveness and convergence. While the predominantly positive affective image across most provinces indicates a general satisfaction with high-quality destinations, the underlying cognitive patterns show significant variations that warrant deeper discussion [27].

For instance, Guizhou’s relatively lower affective image, despite its rich cultural and natural assets, highlights a potential gap between destination resource quality and the perceived service experience. This finding aligns with a body of literature arguing that tangible service elements are critical in translating destination attributes into positive tourist emotions [28]. Our result thus confirms that even for resource-rich destinations, perceived service deficiencies can significantly dampen affective responses [29]. Conversely, Shanghai’s distinct urban tourism experience is consistent with established research on mega-city tourism that identifies modernity, commercial culture, and multiculturalism as key perceptual drivers [30,31]. Our study extends this literature by quantitatively demonstrating how this unique cognitive profile creates a distinct overall image at a national comparative scale, reinforcing the idea that a well-rounded combination of resources, rather than a single core attraction, enhances destination image [32].

Perhaps most intriguingly, the similarities observed between geographically distant regions, such as Hainan and Northeast China, challenge a purely static, geography-based view of image formation. This pattern can be largely attributed to the well-documented phenomenon of seasonal migration in China [33], a trend also fueled by developments in niche markets like ice–snow tourism [34]. Our finding contributes a new dimension to this research by showing that these mobility patterns do not merely reflect travel behavior but actively construct a shared cognitive image across disparate regions. This suggests that social networks and tourist flows are powerful, dynamic forces in shaping destination perception, a notion that echoes recent calls for a more relational understanding of place branding [35].

4.2. Deconstructing the Cognitive–Affective Linkage: Insights from Machine Learning

Our use of random forest regression offers a nuanced view of the intricate relationship between cognitive and affective images, a central theme in tourism image research [36,37]. The model’s relatively low R² value, while indicating the inherent unpredictability of human emotion, is a finding consistent with other studies in computational social science that attempt to model sentiment from complex, user-generated textual data [38,39]. As such, the model’s primary value lies not in its predictive precision but in its explanatory power to identify significant influencing factors.

One of the key insights is the high feature importance of highly specific, iconic cultural symbols. The prominence of a local snack in shaping a destination’s affective image, for example, provides strong empirical support for the growing body of literature on gastronomy tourism and symbolic consumption [40]. Our finding extends this research by demonstrating that the cognitive–affective link is not just driven by general attributes but can be powerfully anchored to micro-level, tangible cultural icons. This suggests that symbolic consumption plays a more granular and emotionally resonant role in destination image formation than previously understood.

Furthermore, the significant negative impact of operational keywords related to service friction (e.g., queuing) on affective scores confirms a large body of service management literature identifying such friction points as key drivers of customer dissatisfaction [41,42]. The primary contribution of our study here is methodological: we demonstrate the use of a machine learning approach to quantify these effects from unstructured user-generated data at a large scale. This offers a new, scalable method for continuously monitoring and diagnosing sources of negative affect in the tourist experience, moving beyond traditional survey-based methods.

4.3. Implications for Tourism Image Theory

Collectively, our findings have several important implications for the theoretical understanding of tourism image. First, by showing how inter-provincial destinations can create image homogeneity, we propose that destination image models should incorporate network effects, moving beyond the traditional focus on a destination’s isolated, intrinsic attributes. This calls for a more dynamic and relational perspective on how regional image systems are formed and evolve.

Second, our machine learning approach reveals the granular nature of the cognitive–affective link [37]. This suggests that the theory could be refined by developing a typology of cognitive cues and examining their differential impacts on emotion. For instance, the emotional resonance of an iconic cultural symbol may function very differently from that of a functional service attribute.

Conceptually, our study contributes by framing destination image not as a static property of a place but as a relational and emergent outcome of tourist–destination interactions, inter-regional mobility, and the symbolic meanings attached to micro-level cultural elements. This perspective enriches the classic cognitive–affective framework [28] by adding layers of social network dynamics and symbolic interactionism.

5. Recommendations for Inter-Provincial Tourism Management and Development Under the Similarity–Difference Matrix

Comprehensive image similarity measures the degree of resemblance between two provinces as high-quality tourism destinations in terms of tourists’ affective, cognitive, and overall perceptions, as well as the cognitive influence on affective image perception. It reflects the commonalities between pairs of provinces. Network edge weights gauge the strength of connections between two provinces within the network, indicating their level of interaction and significance in the global network context. Integrating the two metrics balances local similarity and global network structure, providing a comprehensive analytical perspective. This integration not only advances the scientific and practical value of the research but also offers robust support for tourism collaboration and resource optimization.

Category I (High Similarity–High Edge Weight, High Similarity–Medium Edge Weight, Medium Similarity–High Edge Weight): Strong Synergistic Relationship, Suitable for Deep Collaboration

This relationship indicates significant commonalities and a solid foundation for close cooperation between the provinces. For instance, Beijing and Shaanxi, as cultural and historical tourism destinations, can deepen their collaboration through joint promotion of cultural heritage and the development of cross-provincial tourism routes, thereby enhancing their appeal to international tourists. Specific strategies include jointly building an international tourism brand, optimizing cross-provincial tourism service experience, and developing in-depth cultural tourism products.

Category II (Medium Similarity–Medium Edge Weight): Moderate Synergistic Relationship, Suitable for Gradual Promotion of Collaboration

This relationship suggests potential for cooperation, though it requires further development. For example, Shaanxi and Sichuan could gradually strengthen their collaboration through joint marketing initiatives (e.g., culinary tours) and resource sharing. Specific strategies include highlighting regional characteristics and launching joint marketing, promoting resource sharing and improving tourist mobility, and seeking policy support and improving infrastructure.

Category III (Medium Similarity–Low Edge Weight, High Similarity–Low Edge Weight): Potential Synergistic Relationship, Requiring Enhanced Collaboration

These inter-provincial relationships, despite their high comprehensive similarity, exhibit limited connectivity with other provinces within the network. This relationship indicates significant resource complementary between them, rather than direct competition. For instance, Fujian (coastal) and Jiangxi (mountainous) could jointly develop themed tourism routes, such as “Mountain–Sea Journeys,” leveraging their respective resource advantages to attract tourists. Specific strategies include developing complementary themed tourism routes, innovating collaboration models to enhance appeal, and promoting public–private partnerships to ensure project sustainability.

Category IV (Low Similarity–Medium Edge Weight, Low Similarity–High Edge Weight): Complementary Relationship, Suitable for Differentiated Collaboration

These provinces exhibit low comprehensive similarity but high edge weights, indicating untapped collaborative potential. For example, Jiangsu and Zhejiang could gradually strengthen their cooperation through joint promotional activities and pilot projects. Specific strategies include enhancing interaction and launching joint promotions, initiating pilot projects to explore new models, and leveraging existing tourist resources to extend travel experiences.

Category V (Low Similarity–Low Edge Weight): Independent Relationship, Suitable for Autonomous Development

This relationship suggests that the tourism markets of these provinces are relatively independent, with limited potential for collaboration. However, each province possesses unique tourism resources and development potential. For instance, Yunnan (adventure tourism) and Guangxi (wellness tourism) can focus on developing niche tourism markets with differentiated positioning. Specific strategies include deepening niche markets and building unique brands, establishing initial collaborative links and exploring potential synergies, and formulating long-term development plans to gradually enhance competitiveness.

These strategies not only promote the collaborative development of inter-provincial tourism destinations but also optimize regional tourism resources, leading to higher development levels. By tailoring recommendations to the unique characteristics of each category, this study provides a nuanced framework for enhancing tourism efficiency and fostering sustainable growth.

6. Conclusions

This study offers an innovative approach to constructing a multidimensional image perception model, integrating cognitive, affective, and overall image dimensions. By applying this model to create an evaluation framework for inter-provincial tourism destination similarities and differences, this research addresses an underexplored area in the literature concerning large-scale regional image perception.

First, the study systematically analyzes the image perception of high-quality tourism destinations in China. Our findings revealed a complex perceptual landscape: while most provinces enjoy a predominantly positive affective image, significant variations exist. The cognitive and overall images demonstrated high degrees of uniqueness, while also revealing patterns of homogeneity driven by factors like geographical proximity and tourist mobility.

Second, by exploring the cognitive–affective relationship, our analysis identified the specific attributes most influential in shaping tourist emotions. The findings highlight the significant impact of not only iconic attractions and general scenic terms but also micro-level cultural symbols like local foods and crucial service friction points.

Finally, the study successfully developed a novel matrix. This analytical tool, integrating both static similarity and dynamic network connections, classified inter-provincial relationships into five distinct categories, revealing that most provinces exhibit weaker or more complex relationships.

From a theoretical perspective, this study contributes by extending the traditional cognitive–affective theory with a more dynamic and relational viewpoint, demonstrating how network effects shape destination images. It also refines the theory by revealing the granular and symbolic nature of the cognitive–affective link, highlighting the powerful role of specific cultural icons in emotional formation.

From a practical perspective, this study provides several actionable insights. For example, our findings suggest that destination managers in provinces with similar cognitive but different affective images (a distinction made possible by our matrix) should prioritize emotion-focused storytelling in their branding. Furthermore, the identification of specific service friction points like ‘queuing’ provides clear, data-driven priorities for operational improvements to enhance the tourist experience. The similarity–difference matrix itself serves as a strategic tool for identifying optimal partners for regional tourism collaboration.

7. Limitations and Further Research

This study primarily focuses on the tourism image perception and explores the similarities and differences between high-quality tourism destinations across provinces based on big data. However, there are still areas that could be further improved. From a data perspective, this study relies solely on big data for analysis. Future research could integrate small data and field data, such as surveys and interviews, to further validate and supplement the findings from big data analysis [43]. Methodologically, text processing primarily focuses on word cleaning in this study. Although part-of-speech (POS) tagging was not performed, this omission did not significantly impact the overall analysis results, as the research objectives emphasize statistical features or thematic distributions at the word level rather than in-depth syntactic analysis [44]. Future research could consider incorporating POS tagging techniques to enhance the granularity of text processing [45]. Theoretically, this study only explores the relationship between affective and cognitive images in depth. Future research could quantify the complex network relationships among affective, cognitive, and overall images, building a more detailed theoretical framework [46]. Building on these points, future inquiry could also expand by conducting comparative cross-country studies, employing longitudinal analyses to track image evolution, and testing other advanced machine learning models, all of which would contribute to a deeper and more dynamic understanding of tourism image perception.