Spatiotemporal Distribution Heterogeneity and Nonlinear Driving Factors of Accommodation Establishments in Xinjiang: An XGBoost–SHAP Approach

Abstract

Accommodation establishments constitute a core component of tourism infrastructure, and their location choices directly affect water resource utilization, land pressure, and the spatial equilibrium of tourism development—issues that are particularly acute in vast arid regions. Yet the spatial organization of accommodation supply across extensive drylands characterized by fragmented oasis distribution, and the reasons why standard and non-standard accommodation follow divergent location logics, remain poorly understood. This study addresses three questions: (1) How are nine accommodation categories, differentiated by type and quality, distributed across Xinjiang? (2) Do directional spatial associations exist among categories that are consistent with hierarchical, path-dependent development? (3) Which factors drive these patterns, and do their effects exhibit the nonlinearity and threshold behavior predicted by location theory? Drawing on 12,073 accommodation establishments from the Ctrip platform, we construct a staged analytical framework in which each technique answers a specific question: the nearest-neighbor index and standard deviational ellipse characterize global patterns; kernel density estimation and OPTICS clustering identify local agglomerations; directional local co-location quotients measure asymmetric spatial associations; and XGBoost–SHAP isolates nonlinear drivers and threshold effects. Results reveal a highly concentrated “single-core, multi-center” structure anchored by Urumqi, Yining, and Kashgar, with rapid expansion toward the Ili Valley, Kashgar, and Altay since 2019. Standard accommodation tracks urban centrality and transport nodes, while non-standard accommodation tracks tourism resource endowments, consistent with location-theoretic expectations. Directional co-location analysis reveals hierarchical spatial associations among categories, and driving factors exhibit pronounced nonlinear threshold effects. From a sustainability perspective, the identified thresholds—elevation (1360 m), water-body proximity, and distance to rural tourism demonstration sites (3 km)—constitute quantifiable, spatially explicit sustainability indicators that can be incorporated into planning tools to monitor and steer accommodation development away from ecologically sensitive zones. Global Moran’s I diagnostics of model residuals (reduction of 83–99.7%) suggest that these findings are unlikely to be artifacts of spatial autocorrelation; this diagnostic, however, complements rather than replaces spatially blocked validation. The study contributes category-differentiated, spatially directed evidence for policies balancing tourism expansion against water security and ecosystem integrity, serving sustainable tourism development in arid-region destinations.

1. Introduction

The Xinjiang Uyghur Autonomous Region, China’s largest provincial-level administrative division at approximately 1.66 million km², occupies the heartland of Eurasia and serves as a critical node of the Belt and Road Initiative [1]. Tourism has become a major driver of regional economic transformation, receiving over 300 million visitors and generating RMB 355.2 billion in tourism revenue in 2024 [2]. This expansion occurs in one of the world’s most water-scarce and ecologically fragile environments: precipitation is sparse and unevenly distributed, settlements are confined to discontinuous oasis corridors, and tourism growth competes with agriculture and ecosystems for limited water and land resources [3,4,5]. Accommodation establishments are among the most spatially fixed and resource-intensive components of tourism supply; their location choices determine not only market performance but also the environmental sustainability of destination development [6,7]. Understanding the locational determinants of accommodation and their spatial interaction structure in arid frontier regions is thus both a scientific question and a planning necessity.

Theoretically, the spatial organization of accommodation can be grounded in central place theory and location theory [8,9,10]. Central place theory predicts that service facilities arrange hierarchically according to market thresholds and ranges: higher-order, capital-intensive establishments concentrate in larger urban centers, while lower-order establishments fill peripheral settlements [8]. Location theory further posits that service facilities trade off transport costs, market accessibility, and land rent, producing predictable spatial regularities—standardized, capital-intensive hotels anchor at the thickest markets such as central business districts and transport nodes [8,9,10]. Hotel location research has repeatedly confirmed these regularities, from early analyses of tourist location preferences to formal theories of intra-urban hotel location [9,10]. These classical frameworks anticipate a market hierarchy in which urban centrality, population density, and transport accessibility constitute the primary gravitational fields shaping accommodation location.

Empirical studies of traditional hotels broadly support these expectations. Early point-pattern analyses documented concentrated, center-oriented hotel distributions [8]; historical case studies traced the co-evolution of hotel location with urban structure and tourism development [10]; accessibility-based models and concentric-zone models have been validated in multiple cities [8,9]. Recurring determinants include transport accessibility and commercial agglomeration, with high-grade and mid-to-high-quality hotels concentrating around transport infrastructure and commercial centers [11,12,13,14]. Chinese evidence is illustrative: intra-urban analyses of Urumqi have traced the agglomeration of high-grade hotels along transport corridors and commercial centers [11,12,13,14]. However, most evidence remains at the intra-urban scale [9]: how these regularities operate at regional scales, across discontinuous oases, mountain corridors, and peripheral destinations, is largely unknown.

Agglomeration economies provide a complementary explanation for why accommodation clusters and why their segments differentiate [8]. Co-located hotels benefit from localization economies—specialized labor pools, shared supplier networks, and knowledge spillovers—and urbanization economies—diversified amenities, infrastructure, and visitor flows [8,9]. However, agglomeration returns are uneven across accommodation types. Standard hotels primarily exploit urbanization economies in city centers, whereas non-standard accommodation—homestays, guesthouses, and agritourism lodges—tends to follow resource-oriented location logics in which natural and cultural assets compensate for thinner local markets [15]. The Tourism Area Life Cycle (TALC) further predicts that as destinations evolve from exploration through development to consolidation, the accommodation mix shifts from small-scale, resource-dependent establishments toward larger, market-oriented properties, generating divergent spatial trajectories for the two segments [16,17]. Shared accommodation research further confirms that non-standard accommodation exhibits systematically different spatial dynamics across the destination life cycle compared with standard hotels [18]. Together, these theories imply distinct spatial signatures: standard accommodation tracks urban centrality, while non-standard accommodation tracks tourism resource endowments [16,17,19].

Consistent with this expectation, research on non-standard accommodation—particularly homestays and Airbnb listings—has documented agglomeration around scenic areas, transport infrastructure, and culturally distinctive sites. Studies of rural homestays around Beijing identified significant clustering around scenic spots and transport hubs [20]; Chongqing homestays exhibit a “single-pole, multi-core” structure driven jointly by tourism resource endowments and urban centrality [21,22,23]; comparisons between star-rated hotels and Airbnb in Shanghai revealed divergent spatial logics between standardized and sharing-economy formats within the same urban system [24,25,26,27]; and the Barcelona case further confirmed that shared accommodation spatial distribution is jointly constrained by tourist attractions and urban morphology, displaying complementary rather than overlapping patterns with traditional hotels [28]. However, evidence from arid regions of western China remains scarce, and whether non-standard accommodation is systematically more resource-dependent than traditional hotels has rarely been tested at regional scales.

A further theoretical consideration concerns functional form. Classical location models typically assume linear or log-linear driver–response relationships, yet both location theory and evolutionary economic geography suggest that spatial processes are inherently nonlinear and path-dependent [19,29]. Market-size thresholds imply that below a critical demand level, investment risk deters entry; above it, agglomeration returns accelerate clustering [19]. In arid regions, water-body proximity may impose a near-deterministic constraint on resort-type development, producing sharp thresholds rather than smooth gradients. Evolutionary economic geography emphasizes that such thresholds originate from historically contingent, self-reinforcing processes: early entry creates spatial lock-in, subsequent entrants face increasing costs of deviation, and small differences in initial conditions can produce large, divergent spatial outcomes [19,30]. Path creation and institutional trigger events in destination evolution further suggest that tourism spatial patterns are not the product of equilibrium processes but the superposition of critical junctures and cumulative causation [29,30]. These arguments imply that linear specifications may misidentify driver–response relationships, necessitating interpretable nonlinear methods to recover the underlying spatial-economic mechanisms.

Methodologically, advances in interpretable machine learning and local spatial analysis provide new tools for accommodation geography. The XGBoost–SHAP framework has been widely applied in tourism geography, land-use analysis, and environmental science [31,32,33,34,35,36]; extensions of the co-location quotient capture localized, asymmetric spatial associations between categorical point patterns [37,38]. What remains scarce is their staged integration in arid-region accommodation systems—chaining description (NNI, SDE, KDE), decomposition (OPTICS, LCLQ), and explanation (XGBoost–SHAP) into a coherent analytical pipeline, systematically implemented across the full matrix of nine accommodation categories × two perspectives.

Synthesizing the above literature and theoretical analysis, this study identifies five critical gaps: (1) Geographic scale—most empirical studies are confined to single cities or provinces, lacking regional-scale analyses capable of capturing the full spatial heterogeneity of vast, internally diverse territories [3,39]; (2) Methodology—linear regression frameworks may produce misleading conclusions by ignoring nonlinear relationships, threshold effects, and complex interactions among drivers [40,41]; (3) Theme—standard and non-standard accommodation are rarely integrated within a unified analytical framework, and category-specific nonlinear driving mechanisms have not been systematically decomposed; (4) Spatial association—asymmetric spatial signals (where the prior establishment of one category may directionally influence the subsequent location decisions of another) remain largely unexplored [42,43]; (5) Sustainability—the ecological thresholds and spatial planning implications of accommodation expansion in arid regions have not been quantified.

To address these gaps, this study poses three research questions: Q1—What spatial distribution patterns do nine accommodation categories exhibit in Xinjiang, and how do their agglomeration–dispersion characteristics differentiate by type and quality? Q2—Do directional spatial associations exist among categories, and is their hierarchical structure consistent with central place theory and path-dependence expectations? Q3—Which drivers shape these patterns, do their effects exhibit nonlinear thresholds, and can these thresholds provide spatially explicit reference points for sustainable tourism planning?

2. Materials and Methods

2.1. Study Area

Xinjiang Uyghur Autonomous Region is located in the Eurasian hinterland (73°32′–96°21′ E, 34°22′–49°33′ N) and covers approximately 1.66 million km². Its geomorphology is often described as ‘three mountains enclosing two basins’, comprising the Altai Mountains, Junggar Basin, Tianshan Mountains, Tarim Basin and Kunlun Mountains (Figure 1). This setup is good at generating high spatial variability in climate, hydrology and ecosystem vulnerability. The Oasis settlements are more or less discontiguous along the piedmont alluvial fans and riverbeds, and in contrast to the Gobi and the desert hinterlands, they limit settlement, mobility and tourism development.

The development of tourism in Xinjiang has been rapid. In 2023, the region received 265.44 million tourist visits and RMB 296.715 billion in tourism revenue, year-on-year increases of 117.04% and 226.93% [44]. In 2024, the number of arrivals had surpassed 300 million, and the spending had reached RMB 359.5 billion. This boom drove the supply-side restructuring, with hotel groups opening hotels in Urumqi, Kashgar, Altay, Yining, and over 600 brand hotels opened by 2024 [45], and the share of star-rated hotels also decreased, leading to diversified and quality-oriented supply.

This shift was supported by policy. The autonomous region added cultural and tourism to the industrial cluster sequence in 2024, making it “nine major industrial clusters” from “eight.” The Three-Year Action Plan for High-Quality Tourism Development (2024–2026) aims to reach 350 million tourists by 2026 and RMB 420 billion, focusing on spatial optimization of accommodation and high-quality hotel clusters. However, the urban-rural imbalance and lack of coordination between standard and non-standard accommodation and mismatch in supply and demand during the different seasons are still not addressed.

For the purposes of accommodation geography, these conditions make Xinjiang a valuable case. The corridors of the Oasis are segments of settlement and tourism within a large-scale desert space with arid conditions [3,4]. Given the multi-ethnic population, there are diverse forms that range from chain hotels to yurt encampments and heritage courtyard guesthouses [46,47]. This change in the star ratings of hotels relative to the growth of chain and informal accommodations represents an institutional shift [5,48].

These conditions—extreme water-resource constraints, oasis fragmentation, multi-ethnic diverse accommodation forms, and the superposition of institutional transition and policy shocks—make Xinjiang a paradigmatic case for sustainable tourism research: the ecological feasibility boundaries of accommodation expansion are most clearly discernible here, and the governance need for category differentiation is most urgent.

2.2. Data Source and Processing

Data on accommodation were obtained from the Ctrip Open Platform. Available online: https://open.ctrip.com (accessed on 17 January 2026), the biggest OTA in China. The keyword queries included all 17 prefecture-level divisions, and the return information included the facility name, coordinates, address, diamond rating, type and opening year. Cleaning included coordinate validation (records outside of 73–96° E and 34–50° N were excluded), deduplication using coordinate hashing (±0.001°) with fuzzy name matching and manual verification of ambiguous names. The final data set has 12,073 valid establishments, 11,857 (98.2%) of which have information for the first year of operation. Physiographic, accessibility and socioeconomic variables were assembled (

Table 1. Data sources and processing methods.

Corresponding Variables	Data Format	Processing Method	Data Source
Elevation; Slope	GeoTIFF	Slope derived using the Spatial Analyst tools in ArcGIS Pro 3.6	Geospatial Data Cloud (https://www.gscloud.cn/)
Fractional vegetation cover	GeoTIFF	Maximum-value compositing	NASA Earthdata (https://www.earthdata.nasa.gov/)
Mean annual temperature; Mean annual precipitation	GeoTIFF	Aggregation of monthly means	National Tibetan Plateau Data Center (https://data.tpdc.ac.cn/)
Distance to rivers/lakes; distance to main roads	Shapefile	Network analysis; Euclidean distance calculation	OpenStreetMap (https://www.openstreetmap.org/)
Urbanization intensity	GeoTIFF	Radiometric calibration; regional statistical aggregation	National Earth System Science Data Center (https://www.geodata.cn/)
Regional GDP; per capita GDP; tertiary-industry share	Excel	Data cleaning; county-level spatial join	County-level statistical bulletins of Xinjiang Uygur Autonomous Region
Distance to railway stations; distance to administrative villages; distance to expressway entrances; distance to airports; distance to parking lots; distance to county seats; catering and shopping POI density; distance to higher education institutions; distance to rural tourism demonstration sites	Shapefile/Excel	Batch geocoding using an API; kernel density estimation; Euclidean distance calculation	Amap (Gaode Maps) (https://ditu.amap.com/)
Population density	GeoTIFF	Resampling to a consistent resolution; zonal statistics	WorldPop (https://hub.worldpop.org/)
Distance to high-level scenic areas; distance to low-level scenic areas	Shapefile/Excel	Geocoding verification; multiple-ring buffer analysis	Xinjiang Department of Culture and Tourism (https://wlt.xinjiang.gov.cn/)

), such as elevation, slope, vegetation cover, temperature, precipitation, distances from the water and from the road, intensity of urbanization, GDP, share of the tertiary sector, accessibility, population density and distances to the areas that were graded for their scenery. All variables were processed through spatial analysis before being incorporated into the models.

A three-tier hierarchical classification was constructed to capture structural heterogeneity. The first tier distinguishes standard accommodation (n = 7450) from non-standard accommodation (n = 4623). Standard accommodation includes hotels (n = 7189), establishments with ‘binguan’ in their names (n = 229), other standard types (n = 30), and resorts or specialty hotels (n = 2). Non-standard accommodation includes bed-and-breakfasts (n = 4202), inns (n = 108), and other types (n = 313), including youth hostels, agritainment (nongjiale) and serviced apartments. Classification used direct type matching, keyword disambiguation for 242 facilities containing ‘lüguan’ and manual review of remaining ambiguous cases.

The second tier captures quality level. Given the decline in star-rated participation and the frequent exclusion of non-standard accommodation from that system, we used Ctrip diamond ratings (0–5 diamonds), which broadly correspond to star ratings. Economy (0–2 diamonds) and comfort (3 diamonds) categories were merged into a mass-market tier (n = 10,947, 90.67%). Upscale (4 diamonds) and luxury (5 diamonds) categories were merged into a premium tier (n = 1126, 9.33%).

The third tier cross-classifies type and quality, yielding nine categories: mass-market standard (MS, n = 6510), mass-market non-standard (MNS, n = 4437), mass-market total (MM, n = 10,947), premium standard (PS, n = 940), premium non-standard (PNS, n = 186), premium total (PM, n = 1126), total standard (ST, n = 7450), total non-standard (NS, n = 4623), and all accommodations (n = 12,073).

2.3. Variable Selection

Based on existing literature and data availability, 23 explanatory variables were organized into six dimensions: natural conditions, socioeconomic context, transport accessibility, population and market conditions, public service accessibility, and tourism resources (

Table 2. Selection and quantification of explanatory variables.

Dimension	Variable	Symbol	Measurement	Mechanism
Natural environment	Elevation	Elev	DEM elevation (m)	Terrain constraint
	Slope	Slope	DEM slope (degree)	Buildability
	Fractional vegetation cover	FVC	NDVI-derived FVC	Ecological amenity
	Distance to rivers/lakes	D_Riv	Nearest water body (km)	Waterfront amenity
	Temperature	MAT	Mean annual temperature	Thermal comfort
	Precipitation	MAP	Mean annual precipitation	Hydro-ecological condition
Socioeconomic context	Urbanization intensity	UI	Nighttime-light intensity	Urban activity
	Regional GDP	RGDP	Spatialized GDP	Economic scale
	Per capita GDP	GDPpc	Spatialized GDP per capita	Consumption capacity
	Tertiary-industry share	PTI	Service-sector share (%)	Service economy
Transport accessibility	Distance to railway station	D_Rail	Nearest railway station (km)	Rail access
	Distance to administrative village	D_AdmVil	Nearest village committee (km)	Rural service base
	Distance to main road	D_MRoad	Nearest main road (km)	Road access
	Distance to expressway entrance	D_ExpEnt	Nearest expressway entrance (km)	Regional access
	Distance to airport	D_Air	Nearest airport (km)	Air access
	Distance to parking lot	D_Park	Nearest parking lot (km)	Self-drive support
Population and market	Distance to county seat	D_County	Nearest county seat (km)	Market/service center
	Catering/shopping POI density	CSP_Den	Kernel density (units/km2)	Commercial support
	Population density	PD	People per 100 m grid	Local demand/labor
Tourism resources	Distance to higher education institution	D_HEI	Nearest HEI (km)	Higher education
	Distance to rural tourism site	D_RTDS	Nearest demonstration site (km)	Rural tourism pull
	Distance to high-level scenic area	D_HLSA	Nearest 4A/5A site (km)	Major scenic pull
	Distance to low-level scenic area	D_LLSA	Nearest ≤ 3A site (km)	General tourism setting

). The table reports each variable, symbol, measurement and expected relationship with accommodation distribution [46,49,50,51,52,53,54].

Variable selection was not purely data-driven but grounded in the operationalization of the theoretical framework outlined above. Specifically: (1) Natural-condition variables (elevation, slope, FVC, temperature, precipitation) correspond to the physical constraint surface in location theory—in arid regions, these variables jointly define the ecological feasibility boundary for human settlement and tourism activity [49,55]; (2) Hydrological proximity (distance to rivers/lakes) constitutes a near-deterministic resource constraint in arid regions, where resort-type development depends on water bodies far more than in humid regions [4]; (3) Transport accessibility variables (distance to railway stations, airports, highway entrances, trunk roads, parking lots) correspond to the transport-cost term in Weberian location theory—standard accommodation, given the time-sensitivity of business travelers, exhibits stronger dependence on transport nodes [10]; (4) Socioeconomic/market variables (urbanization intensity, GDP, GDP per capita, tertiary industry share) correspond to market thresholds and ranges in Christallerian central place theory—urban centrality and economic structure determine the scale and hierarchy of accommodation demand [8]; (5) Public service accessibility (distance to county seats, administrative villages, universities; catering and shopping POI density; population density) corresponds to urbanization economies in agglomeration theory—diversified amenities and labor pools provide externality support for hotel operations [8]; (6) Tourism resource variables (distance to rural tourism demonstration sites, high-grade/low-grade scenic areas) correspond to resource dependence in TALC—non-standard accommodation location logic is expected to track tourism asset distribution more closely [16]. This six-dimensional framework ensures that each variable has an explicit theoretical role rather than entering solely on the basis of statistical correlation.

2.4. Research Methods

2.4.1. Kernel Density Estimation (KDE) and Dual-Perspective Density Framework

Accommodation patterns show two processes: the supply-side location decisions by the investors and the demand-side consumption by guests. Facility density and demand intensity are not necessarily the same, which may cause the location mechanisms to be confused, so a dual-perspective KDE was used to create two spatial variables: spatial supply intensity (SSI) and market heat-weighted density (MHWD) [56,57,58,59]. KDE measures neighborhood density as a way of measuring local clustering:

$$KDE\left(x,y\right)={\displaystyle \frac{1}{nh}}\sum K\left({\displaystyle \frac{d_i}{h}}\right)$$

(1)

The first dependent variable, Spatial Supply Intensity (SSI), captures supply-side location decisions using standard KDE without demand weighting:

$$SSI\left(x,y\right)={\displaystyle \frac{1}{nh}}\sum K\left({\displaystyle \frac{d_i}{h}}\right)$$

(2)

Market Heat-weighted Density (MHWD) assigns weights to each facility, based on the intensity of participation of guests per year. The market heat index is:

$$H_i={\displaystyle \frac{R_i}{\mathrm{m}\mathrm{i}\mathrm{n}\left(T_i,3\right)}}$$

(3)

where $R_i$ is the total number of Ctrip reviews, $T_i$ = 2025 − $Y_i$ + 1 is the number of years the platform has been operating, and $Y_i$ is the year the platform opened, and min($T_i$, 3) is a correction for the platform’s 3-year rolling review window. The surface of the MHWD is:

$$MHWD\left(x,y\right)={\displaystyle \frac{1}{nh}}\sum H_i\cdot K\left({\displaystyle \frac{d_i}{h}}\right)$$

(4)

The reviews on the platform are not necessarily a reflection of total demand and are interpreted as the “market heat” seen by the platform, as review numbers can be driven by platform penetration, merchant strategies, composition of guests at the platform, and platform pricing and promotion. It is used to compare relative spatial patterns in the same platform environment.

2.4.2. Nearest Neighbor Index (NNI)

The Nearest Neighbor Index (NNI) computes the point clustering as well as the random or dispersed nature of the points:

$$R={\displaystyle \frac{{\bar{r}}_1}{{\bar{r}}_E}}=2\bar{D}\sqrt{{\displaystyle \frac{n}{A}}}$$

(5)

where ${\bar{r}}_1$ is the observed mean nearest-neighbor distance, ${\bar{r}}_E$ the expected distance under complete spatial randomness, $\bar{D}$ is the observed mean nearest-neighbor distance, n the number of features, and A the total area. $R$ < 1: Clustering, $R$ = 1: Randomness, $R$ > 1: Dispersion, tested for significance by Z-value. The Average Nearest Neighbor tool in ArcGIS Pro 3.6 was used to calculate NNI for all nine accommodation categories and for the three macro-regions (Northern, Southern and Eastern Xinjiang).

2.4.3. Standard Deviational Ellipse (SDE)

The Standard Deviational Ellipse (SDE) was applied to characterize directional trends and spatial dispersion. It identifies the weighted mean center, orientation and eccentricity of each accommodation category:

$${\bar{X}}_w={\displaystyle \frac{\sum w_i{x}_i}{\sum w_i}}, {\bar{Y}}_w={\displaystyle \frac{\sum w_i{y}_i}{\sum w_i}}$$

(6)

$$\mathrm{t}\mathrm{a}\mathrm{n}\left(2\theta \right)={\displaystyle \frac{\sum w_i^2\left({\bar{x}}_i-{\bar{X}}_w\right)\left({\bar{y}}_i-{\bar{Y}}_w\right)}{\sum w_i^2{\left({\bar{x}}_i-{\bar{X}}_w\right)}^2-\sum w_i^2{\left({\bar{y}}_i-{\bar{Y}}_w\right)}^2}}$$

(7)

where $w_i$ is the weight assigned to point $i$, $x_i$ and $y_i$ are the coordinates, and $\theta$ is the rotation angle suggesting the principal directional axis. The Directional Distribution tool in ArcGIS Pro 3.6 was used to generate one-standard-deviation ellipses, covering approximately 68% of points, for each accommodation category.

2.4.4. OPTICS Clustering

For clustering with varying densities, Ordering Points To Identify the Clustering Structure (OPTICS), an extension of DBSCAN, was used. OPTICS constructs a reachability plot based on core distance (distances to the k-th neighbors, k = MinPts) and reachability distance (maximum of core distance and distance to the neighbor), and thus permits the extraction of clusters of various densities from a single data set. The choice of OPTICS was motivated by the area size of Xinjiang (~1.66 million km²), which creates significant density heterogeneity, so that a single DBSCAN threshold might split urban clusters or fail to identify peripheral clusters. The parameters were calibrated iteratively: MinPts = 100 and $\epsilon$ = 70 km. In the analysis, 13 clusters were found, comprising 11,382 points (94.3%) and 691 points (5.7%) as noise.

To assess the robustness of clustering results to parameter selection, we re-ran clustering within the same tool across a 16-combination parameter grid of MinPts ∈ {50, 80, 100, 150} × search distance ∈ {30, 50, 70, 100} km, and measured point-level assignment agreement of each partition relative to the baseline configuration (MinPts = 100, search distance = 70 km) using the Adjusted Rand Index (ARI) and Normalized Mutual Information (NMI)

2.4.5. Local Colocation Quotient (LCLQ)

The Local Colocation Quotient (LCLQ) was used to quantify spatial association between categories, a point-pattern statistic that quantifies the extent to which one category is attracted to or repelled from another [37,38]. Contrary to global indices, LCLQ goes beyond this and considers the context of each point, which allows for heterogeneous colocation in large, diverse areas [60]. The LCLQ of category to category is:

$$LCLQ\left(A_i\to B\right)={\displaystyle \frac{N_{A_i\to B}/N_B}{\left(N-N_{A_i}\right)/\left(N-1\right)}}$$

(8)

where is the weighted number of points in ‘s neighborhood (temporal precedence), are the total number of points, and N is the total number of points from all categories. LCLQ is asymmetric: LCLQ (A → B) is not the same as LCLQ (B → A), as it represents directional spatial association. This was done in ArcGIS Pro 3.6 using a Gaussian distance-decay kernel, and significance was determined by the 999 Monte Carlo permutations at p < 0.05 [61,62]. This was achieved by excluding facilities open for less than one year, establishing a temporal ordering between prior and subsequent establishments so that the analysis captures sequential spatial association rather than simultaneous co-occurrence [63,64]. Of the nine categories, eight directional associations were analyzed.

It must be emphasized that LCLQ measures spatial association, not causal direction. Temporal precedence is a necessary but insufficient condition for causal inference; it cannot eliminate self-selection bias or omitted-variable bias. The sequential spatial emergence of accommodation categories may be driven by unobserved third-party factors. Accordingly, all LCLQ results in this study should be interpreted as “directional spatial associations”—that is, in areas where Category A facilities are dense, the probability of Category B facilities occurring is significantly higher than random expectation—rather than A “guiding” or “driving” B’s location decisions. Any mechanistic interpretation is advanced as a hypothesis requiring further testing with longitudinal data or quasi-experimental designs.

2.4.6. XGBoost Model

XGBoost (eXtreme Gradient Boosting) is a distributed gradient-boosting algorithm built on the GBDT framework. It minimizes the following objective function:

$$Obj={\displaystyle \sum _{i=1}^n}l\left(y_i,{\widehat{y}}_i\right)+{\displaystyle \sum _{k=1}^K}\Omega \left(f_k\right)$$

(9)

where $n$ is the number of observations; $l\left(y_i,{\widehat{y}}_i\right)$ is the loss function; $K$ is the number of trees; and $\Omega \left(f_k\right)$ is the regularization term penalizing model complexity:

$$\Omega \left(f_k\right)=\gamma T+{\displaystyle \frac{1}{2}}\lambda {\parallel w\parallel }^2$$

(10)

where $T$ is the number of leaves, $w$ denotes leaf weights, and $\gamma$ and $\lambda$ are regularization hyperparameters.

Eighteen XGBoost regression models were constructed (9 categories × 2 density perspectives). The dependent variables were SSI, calculated as standard KDE from coordinates, and MHWD, calculated as demand-weighted KDE using review counts normalized by operational tenure. Data were split 70/30 into training and testing sets through stratified random sampling. Hyperparameters were learning_rate = 0.05, max_depth = 6, n_estimators = 500, subsample = 0.8 and colsample_bytree = 0.8, with early_stopping_rounds = 20. Model performance was evaluated using $R^2$, RMSE and MAE.

To provide a complementary diagnostic of residual spatial structure, we computed global Moran’s I for all 18 models. Specifically, a k-nearest-neighbor spatial weight matrix was constructed from the longitude–latitude coordinates of each accommodation point. Global Moran’s I was computed for both the dependent variable and model residuals, with significance assessed via 999 random permutation tests. Residual Moran’s I close to the spatial-randomness expectation indicates limited remaining spatial autocorrelation after fitting. However, this diagnostic does not assess out-of-area transferability or by itself rule out optimistic fit arising from a random split; it is therefore interpreted alongside, rather than as a substitute for, spatially blocked validation. This use of a residual spatial autocorrelation diagnostic follows Li et al. [55].

2.4.7. SHAP Interpretability Analysis

SHAP values were used to interpret XGBoost outputs by estimating each feature’s contribution to model predictions under a cooperative game-theory framework. The SHAP value for each feature is:

$${\varphi }_j=\sum _{S\subseteq N\backslash \left\{j\right\}}{\displaystyle \frac{\left|S\right|!\left(\left|N\right|-\left|S\right|-1\right)!}{\left|N\right|!}}\left[f_{S\cup \left\{j\right\}}\left(x_{S\cup \left\{j\right\}}\right)-f_S\left(x_S\right)\right]$$

(11)

where $N$ is the full feature set; $S$ is a subset excluding $j$; $\left|S\right|$ is the cardinality; and $f_S\left(x_S\right)$ is the prediction using only subset $S$.

The interpretation included three parts: (1) global feature importance based on mean absolute SHAP values ($\left|{\varphi }_j\right|$); (2) partial dependence profiles using LOWESS-fitted SHAP-feature scatter plots; and (3) interaction effects based on TreeSHAP pairwise interaction values:

$$I\left(i,j\right)=E\left[f\left(x\right)|x_i,x_j\right]-E\left[f\left(x\right)|x_i\right]-E\left[f\left(x\right)|x_j\right]+E\left[f\left(x\right)\right]$$

(12)

Interaction direction and magnitude were used to classify three interaction types: amplifying, compensating and threshold interactions. Figure 2 gives the analytical framework.

3. Results

3.1. Spatial Distribution Patterns

3.1.1. Distribution Characteristics of Accommodation Types

As of 2025, Xinjiang contained 12,073 accommodation establishments. Mass-market facilities accounted for 10,947 establishments (90.7%), while premium facilities accounted for 1126 (9.3%). Standard accommodation totaled 7450 establishments (61.7%), compared with 4623 non-standard establishments (38.3%). Northern Xinjiang contained the largest share (7406, 61.3%), followed by Southern Xinjiang (3373, 27.9%) and Eastern Xinjiang (712, 5.9%). Within Northern Xinjiang, Ili, Urumqi, Altay and Changji together accounted for 87.3% of the regional total. In Southern Xinjiang, Kashgar was the leading prefecture, followed by Bayingolin, Aksu, Hotan and Kizilsu. Eastern Xinjiang was more evenly divided between Turpan and Hami.

Table 3. Number of categorized accommodation types by region in Xinjiang.

Region	Mass-Market Standard	Mass-Market Non-Standard	Mass-Market Total	Premium Standard	Premium Non-Standard	Premium Total	Standard Total	Non-Standard Total	Total
Xinjiang	6510	4437	10,947	940	186	1126	7450	4623	12,073
Northern Xinjiang	3387	3284	6671	574	161	735	3961	3445	7406
Eastern Xinjiang	511	144	655	49	8	57	560	152	712
Southern Xinjiang	2159	924	3083	273	17	290	2432	941	3373
Aksu Prefecture	486	135	621	56	1	57	542	136	678
Altay Prefecture	308	739	1047	81	63	144	389	802	1191
Bayingol Mongol Autonomous Prefecture	607	140	747	72	0	72	679	140	819
Bortala Mongol Autonomous Prefecture	214	126	340	36	5	41	250	131	381
Changji Hui Autonomous Prefecture	402	133	535	39	0	39	441	133	574
Hami City	255	65	320	23	0	23	278	65	343
Hotan Prefecture	225	31	256	32	2	34	257	33	290
Kashgar Prefecture	769	566	1335	107	11	118	876	577	1453
Karamay City	212	34	246	30	0	30	242	34	276
Kizilsu Kirghiz Autonomous Prefecture	72	52	124	6	3	9	78	55	133
Tacheng Prefecture	229	53	282	10	0	10	239	53	292
Turpan City	256	79	335	26	8	34	282	87	369
Urumqi City	1159	453	1612	215	0	215	1374	453	1827
Ili Kazakh Autonomous Prefecture	863	1746	2609	163	93	256	1026	1839	2865

summarizes the categorized distribution.

Standard accommodation had a strong urban character in mass-market standard accommodation (Figure 3a and Figure 4a). There are 6510 establishments, with Urumqi leading the way (1159, 17.8%), followed by Ili, Kashgar, Bayingolin, Aksu and Altay, and the top 15 cities accounting for over 70%. KDE identified one main core and two secondary ones, namely Urumqi as the main core, and Kashgar as well as Yining as the secondary cores. The railway and expressway in Northern Xinjiang, Urumqi-Shihezi-Kuitun, is manifested in a linear high-density belt.

Mass-market non-standard accommodation was commented on in relation to scenic resources (Figure 3b and Figure 4b). Of the 4437 establishments, Ili led with 1746 establishments (39.4%), with the next three being Altay, Kashgar and Urumqi. KDE suggested that the multiple cores existed in Burqin, Kashgar and Tashkurgan, and the multi-poles were located in Urumqi and the areas around Yining and Ili River Valley. The Burqin concentration is indicative of the Kanas Scenic Area.

Mass-market accommodation as a whole had a single-pole, dual-core pattern that aligns with the mountain-basin geography of Xinjiang (Figure 3c and Figure 4c). Ili took the lead among 10,947 establishments, followed by Urumqi, Kashgar and Altay. Urumqi welcomes business guests, Yining offers Ili Valley ecological tourism, and Kashgar provides Silk Road cultural tourism.

The focus of premium standard accommodation was on the primate city (Figure 3d and Figure 4d). Among 940 establishments, Urumqi held 215 (22.9%), including 29 five-star and 186 four-star hotels, followed by Ili (163), Kashgar (107), Altay (81), Bayingolin (72) and Aksu (56). KDE highlighted one major core (Urumqi) and minor cores (Kashgar, Yining, Korla and Burqin) with high investment thresholds and preferences of investors for commercially validated districts.

High-quality natural landscapes, and of course the distinctive cultural resources (Figure 3e and Figure 4e), were particularly important for premium non-standard accommodation. Ili had 93 (50.0%), and Altay had 63 out of 186 establishments. The poles of KDE were located near Kanas Scenic Area and around Nalati Grassland in Xinyuan County. The lack of it in Bayingolin implies that the boutique homestay development is less strong.

Overall, premium accommodation included two models: hotel-dependent model in Urumqi and boutique homestay-led model in Burqin (Figure 3f and Figure 4f). Among 1126 establishments, Ili led (256), followed by Urumqi (215, 19.1%), Altay (144) and Kashgar (118). There were no premium non-standard establishments in Urumqi, while in Burqin, there were 49, putting Burqin first among counties, and suggesting different visitor structures and market logics.

The standard accommodation was in line with the urban hierarchy (Figure 3g and Figure 4g). Among 7450 establishments, Urumqi (1374), Ili (1026), Kashgar (876), Bayingolin (679) and Aksu (542) formed the major agglomerations. Only Urumqi reached a high level; Yining and Kashgar were relatively high.

Outward expansion into new tourism spaces was evident for non-standard accommodation (Figure 3h and Figure 4h). Among 4623 establishments, Ili held 1839 (39.8%), followed by Altay (802), Kashgar (577) and Urumqi (453). In 2023–2025, Tashkurgan saw a net increase of 202 establishments, while 10 homestays were found in the high-altitude border area of Saitula Town, Pishan County.

The 12,073 establishments are single-pole, multiple-core, regardless of their type (Figure 3i and Figure 4i). Ili accounted for 2865 (23.7%), Urumqi 1827 (15.1%), Kashgar 1453 (12.0%) and Altay 1191 (9.9%). At the core of the three-tiered system, related to oasis geography, there was only a high-level core, namely Urumqi, and regional sub-cores were Yining and Kashgar.

3.1.2. Nearest Neighbor Index Analysis

For Xinjiang, the clustering was found to be strong with the overall R value of 0.0675 (

Table 4. Nearest Neighbor Index of categorized accommodation types in Xinjiang.

Region	Mass-Market Standard	Mass-Market Non-Standard	Mass-Market Total	Premium Standard	Premium Non-Standard	Premium Total	Standard Total	Non-Standard Total	Total
Xinjiang	0.067 ***	0.1183 ***	0.0678 ***	0.0909 ***	0.1308 ***	0.0929 ***	0.0653 ***	0.1191 ***	0.0675 ***
Northern Xinjiang	0.0924 ***	0.1206 ***	0.0915 ***	0.1171 ***	0.1195 ***	0.1098 ***	0.0898 ***	0.12 ***	0.0918 ***
Eastern Xinjiang	0.0703 ***	0.1651 ***	0.0679 ***	0.1256 ***	0.1553 ***	0.1196 ***	0.0677 ***	0.1608 ***	0.0658 ***
Southern Xinjiang	0.0554 ***	0.1409 ***	0.0604 ***	0.0759 ***	0.4694 ***	0.091 ***	0.0537 ***	0.1459 ***	0.0592 ***
Aksu Prefecture	0.1069 ***	0.2141 ***	0.0894 ***	0.1249 ***	-	0.1248 ***	0.1025 ***	0.2133 ***	0.0869 ***
Altay Prefecture	0.0678 ***	0.0778 ***	0.0731 ***	0.1477 ***	0.1542 ***	0.0874 ***	0.0634 ***	0.0754 ***	0.069 ***
Bayingol Mongol Autonomous Prefecture	0.0281 ***	0.1989 ***	0.0581 ***	0.0446 ***	-	0.0446 ***	0.0298 ***	0.1989 ***	0.0585 ***
Bortala Mongol Autonomous Prefecture	0.0657 ***	0.2229 ***	0.1036 ***	0.2297 ***	0.2833 **	0.2252 ***	0.0696 ***	0.2203 ***	0.106 ***
Changji Hui Autonomous Prefecture	0.1105 ***	0.1877 ***	0.1096 ***	0.0495 ***	-	0.0495 ***	0.1101 ***	0.1877 ***	0.1085 ***
Hami City	0.0705 ***	0.1957 ***	0.0651 ***	0.1901 ***	-	0.1901 ***	0.0689 ***	0.1957 ***	0.0636 ***
Hotan Prefecture	0.0583 ***	0.1661 ***	0.0627 ***	0.0325 ***	0.89340.773	0.1068 ***	0.0546 ***	0.2367 ***	0.0621 ***
Kashgar Prefecture	0.0519 ***	0.1093 ***	0.0587 ***	0.1145 ***	0.3031 ***	0.1132 ***	0.0492 ***	0.108 ***	0.0565 ***
Karamay City	0.0671 ***	0.0958 ***	0.0712 ***	0.1398 ***	-	0.1398 ***	0.0915 ***	0.0958 ***	0.0923 ***
Kizilsu Kirghiz Autonomous Prefecture	0.1858 ***	0.5185 ***	0.0883 ***	0.2476 ***	0.1517 **	0.2897 ***	0.1784 ***	0.4973 ***	0.0818 ***
Tacheng Prefecture	0.1134 ***	0.3332 ***	0.0979 ***	0.3035 ***	-	0.3035 ***	0.1361 ***	0.3332 ***	0.1005 ***
Turpan City	0.0729 ***	0.1356 ***	0.0751 ***	0.0446 ***	0.2738 ***	0.0463 ***	0.0689 ***	0.1295 ***	0.0724 ***
Urumqi City	0.1205 ***	0.1932 ***	0.1455 ***	0.1703 ***	-	0.1703 ***	0.121 ***	0.1932 ***	0.1479 ***
Ili Kazakh Autonomous Prefecture	0.0732 ***	0.0616 ***	0.0589 ***	0.0548 ***	0.0844 ***	0.067 ***	0.0637 ***	0.0632 ***	0.0604 ***

Note: *** and ** indicate significance at the 1% and 5% levels, respectively.

; p < 0.001). All R values were < 1 and significant. Standard accommodation (R = 0.0653) was more clustered than non-standard (R = 0.1191), and mass-market (R = 0.0678) more than premium (R = 0.0929). The non-standard (R = 0.1183) mass-market was more dispersed than the standard (R = 0.0670) mass-market. Within the overall clustered pattern, premium non-standard accommodation had the highest R value (R = 0.1308), indicating the weakest clustering among all categories.

Among the regions, Southern Xinjiang had the lowest R = 0.0592, the strongest clustering, which was due to the isolation of oasis-islands in the Taklamakan Desert, and Northern Xinjiang had R = 0.0918 and Eastern R = 0.0658. Kashgar had the highest concentration (R = 0.0565), and Urumqi had the lowest (R = 0.1479) in line with the polycentric structure.

3.1.3. Standard Deviational Ellipse and Centroid Migration

Spatial expansion was seen in all categories (

Table 5. Standard deviational ellipse parameters of accommodation facilities in Xinjiang.

Category	Year	N	Longitude (°E)	Latitude (°N)	Long Axis (°)	Short Axis (°)	Azimuth (°)	Eccentricity	Area (km2)
Mass-market standard	2010	423	85.38	43.28	4.12	1.68	69.93	0.9132	195,381
	2015	1401	85.51	43.32	4.08	1.73	71.84	0.9063	199,017
	2020	3140	85.08	43.17	4.34	1.83	69.22	0.9065	225,498
	2025	6412	84.33	42.82	4.76	1.87	68.37	0.9196	253,967
Mass-market non-standard	2010	35	86.15	44.37	3.02	2.31	62.81	0.6443	193,210
	2015	98	85.44	44.41	4.11	2.13	62.05	0.855	242,825
	2020	631	84.93	44.39	3.93	1.93	58.88	0.8707	210,675
	2025	4356	83.36	43.56	4.78	1.86	61.07	0.9214	250,068
Mass-market total	2010	458	85.44	43.36	4.05	1.75	69.52	0.9024	200,000
	2015	1499	85.50	43.39	4.08	1.78	71.19	0.8999	205,081
	2020	3771	85.05	43.37	4.27	1.92	67.66	0.8936	231,100
	2025	10,768	83.94	43.12	4.77	1.96	65.85	0.9118	264,930
Premium standard	2010	63	86.15	43.54	3.13	1.78	66.87	0.8219	156,898
	2015	118	85.27	43.33	4	1.74	69.77	0.9006	197,010
	2020	240	85.28	43.24	3.93	1.81	68	0.887	201,736
	2025	905	84.21	43.06	4.74	1.81	65.54	0.9245	242,788
Premium non-standard	2020	5	86.55	47.39	2.44	0.18	32.67	0.9972	11,606
	2025	184	83.91	44.66	4.39	1.44	51.18	0.9445	175,158
Premium total	2010	63	86.15	43.54	3.13	1.78	66.87	0.8219	156,898
	2015	118	85.27	43.33	4	1.74	69.77	0.9006	197,010
	2020	245	85.30	43.33	3.92	1.87	66.84	0.8792	206,813
	2025	1089	84.16	43.33	4.67	1.89	63.07	0.9147	248,767
Standard total	2010	486	85.48	43.31	4.01	1.69	69.74	0.9067	191,905
	2015	1519	85.49	43.32	4.08	1.73	71.7	0.9058	198,938
	2020	3380	85.09	43.18	4.31	1.83	69.15	0.9053	223,898
	2025	7317	84.32	42.85	4.76	1.87	68.02	0.9199	253,044
Non-standard total	2010	35	86.15	44.37	3.02	2.31	62.81	0.6443	193,210
	2015	98	85.44	44.41	4.11	2.13	62.05	0.855	242,825
	2020	636	84.95	44.41	3.93	1.93	58.6	0.8704	210,564
	2025	4540	83.38	43.60	4.77	1.85	60.64	0.9213	248,567
All	2010	521	85.53	43.38	3.96	1.75	69.38	0.8969	195,853
	2015	1617	85.48	43.39	4.08	1.78	71.09	0.8999	204,530
	2020	4016	85.07	43.37	4.25	1.91	67.63	0.8929	229,688
	2025	11,857	83.96	43.14	4.76	1.95	65.6	0.912	263,612

). Mass-market standard accommodation expanded from 195,381.30 to 253,967.48 km² (+30.0%), its center shifting from 85.38° E, 43.28° N to 84.33° E, 42.82° N (~99.4 km). Mass-market non-standard expanded from 193,210.24 to 250,067.52 km² (+29.4%), with its center shifting from 86.15° E, 44.37° N to 83.36° E, 43.56° N (~240.2 km, the largest displacement), reflecting post-2015 homestay reshaping in the Ili Valley and Altay.

Premium standard accommodation showed the fastest expansion, from 156,898.08 to 242,787.64 km² (+54.7%). In 2010 and 2015, the centroid of its frequency of occurrence was at Urumqi, but in 2025, the centroid moved westward to 84.21° E and 43.06° N (~176.2 km). However, premium non-standard development was different, with the number of premium non-standard establishments increasing from 5, focused on Kanas in 2020, to 184 in 2025, moving the focus south to 83.91° E, 44.66° N, and signifying diffusion from Kanas toward the Ili Valley.

In 2025, the mass-market non-standard centroid was ~0.74° north of the standard centroid, with a displacement of 240.2 km, which was about 2.4 times that of standard accommodation, 99.4 km (Table 5; Figure 5). Non-standard housing grew at a faster rate in areas with more resources, while standard housing was diffused along the urban hierarchy.

3.1.4. Interannual Growth Analysis

Based on 11,857 establishments with valid opening-year records, cumulative growth followed several stages (Figure 6;

Table 6. Development stages of accommodation facilities in Xinjiang (1950–2025).

Stage	Period	Start (Count)	End (Count)	Net Increase
Embryonic	1950–1999	1	46	45
Initial	2000–2009	46	369	323
Rapid Development	2010–2018	369	2809	2440
Explosive Growth	2019–2025	2809	11,857	9048

). Starting from one establishment in 1950, the total reached 46 by 1999, 521 by 2010, 1617 by 2015 and 2809 by 2018. In 2019, the number increased to 3547 and then accelerated to 6778 in 2023, 9448 in 2024 and 11,857 in 2025.

The 75-year trajectory can be divided into four stages (Table 6): embryonic growth from 1950 to 1999, with a net increase of 45; initial growth from 2000 to 2009, with a net increase of 323, or approximately 35 per year; rapid development from 2010 to 2018, with a net increase of 2440, or approximately 286 per year; and accelerated growth from 2019 to 2025, with a net increase of 9048, or approximately 1385 per year. The final stage was 4.8 times larger than the rapid-development period.

Annual additions increased from 508 in 2018 to 738 in 2019 (+45.3%). They then declined to 469 in 2020 owing to the COVID-19 pandemic, before recovering to 696 in 2021, 623 in 2022, 1443 in 2023 and a peak of 2670 in 2024. The acceleration from 2023 to 2024 reflected domestic tourism recovery, social media dissemination and accommodation supply shortages.

During 2023–2025, 6522 establishments were added. Ili ranked first (1911, 29.3%), concentrated in Yining City (587), Xinyuan County (343), Tekes County (261), Gongliu County (158) and Huocheng County (147). Kashgar came second (988, 15.1%), with additions mostly placed in Kashgar City (579) and Tashkurgan (277). Urumqi ranked third (743, 11.4%). The proportion of non-standard additions in Altay (681, 10.4%) was the fourth largest, and there were no standard additions in Burqin County. These 4 prefectures accounted for the increase in establishments by 4323 (66.3%).

3.2. Analysis of Clustering Results

Clustering was performed on the 12,073 accommodation establishments in Xinjiang, using OPTICS clustering. A number of iterations later, the best choice of parameters was determined to be: MinPts = 100 and ε = 70 km. A total of 13 spatial clusters were identified in the analysis, with 11,382 establishments (94.3%) and 691 establishments (5.7%) as noise points (

Table 7. Clustering results of accommodation points in Xinjiang.

Cluster Number	Cluster Name	Main Coverage	Number of Accommodation Points
1	Pamir Cluster	Akto County, Taxkorgan Tajik Autonomous County (along their border)	363
2	Kashgar Urban Cluster	Kashgar City, Shufu County, Shule County	956
3	Yarkant Cluster	Yarkant (Shache) County, Zepu County, northern Yecheng County, western Markit County	187
4	Aksu–Aral Cluster	Aral City, Aksu City, Wensu County, Wushi County, Kalpin County	442
5	Kuqa–Xinhe Cluster	western Kuqa City, eastern Xinhe County, northern Xayar County	337
6	Ili–Bortala Cluster	Ili River Valley, Bortala Mongol Autonomous Prefecture	3163
7	Yanqi Basin Cluster	Korla City, Tiemenguan City, Yanqi Hui Autonomous County, Bohu County, eastern Hoxud County	558
8	Northern Slope of Tianshan Mountains Urban Cluster	Urumqi City, central & eastern Changji Hui Autonomous Prefecture, Karamay City, Shihezi City, Usu City, Toksun County, Gaochang District, Shanshan County, Turpan City	3579
9	Tiemenguan–Ruoqiang Cluster	Tiemenguan City, Ruoqiang County	109
10	Hami Cluster	Yizhou District of Hami City, Xinxing City, southeastern Barkol Kazakh Autonomous County	308
11	Tacheng Basin Cluster	Baiyang City, Tacheng City, Emin County, Yumin County, northwestern Toli County	138
12	Altay Cluster	Fuhai County, Beitun City, Altay City, Burqin County, Habahe County	1065
13	Fuyun–Qinghe Cluster	northern Fuyun County, northern Qinghe County	177

).

To assess the robustness of clustering results to parameter selection, we tested a 4 × 4 parameter grid (MinPts ∈ {50, 80, 100, 150} × search distance ∈ {30, 50, 70, 100} km), measuring point-level assignment agreement against the baseline using ARI and NMI (

Table 8. OPTICS parameter sensitivity analysis.

MinPts	Search Distance (km)	No. of Clusters	Noise (%)	ARI vs. Baseline	NMI vs. Baseline
50	30	54	21.8	0.18	0.61
50	50	53	14.9	0.24	0.66
50	70	13	3.7	0.5	0.8
50	100	15	2.7	0.58	0.84
80	30	38	20.5	0.24	0.63
80	50	24	10.6	0.53	0.77
80	70	14	4.2	0.59	0.87
80	100	9	10.6	0.32	0.64
100	30	26	21.4	0.34	0.66
100	50	18	10.7	0.75	0.84
100	70	13	5.7	1	1
100	100	13	4	0.99	0.97
150	30	21	28.3	0.26	0.62
150	50	16	14	0.72	0.81
150	70	12	8.4	0.98	0.96
150	100	11	7.1	0.98	0.96

).

The sensitivity analysis reveals the following key findings. First, the baseline is highly robust to MinPts: holding search distance = 70 km, increasing MinPts from 100 to 150 produces highly consistent partitions (ARI = 0.98, NMI = 0.96), with the urban-group structure remaining unchanged. Second, search distance = 70–100 km × MinPts = 100–150 constitutes a high-stability platform (ARI ≥ 0.98), indicating that the core cluster structure is insensitive to parameter perturbation within this range. Third, at search distance = 30 km, the distribution is over-fragmented into 21–54 sub-clusters with noise proportions of 21–28%; at 50 km, cluster numbers range from 16 to 53 (depending on MinPts) with 11–15% noise, still excessively fragmented. Fourth, MinPts = 50–80 at search distance = 70 km produces similar cluster counts (13–14) but with reduced noise (3.7–4.2%) and ARI of only 0.50–0.59, indicating that too-small MinPts alters point-level cluster assignments, releasing marginal points that should belong to urban groups into noise or micro-clusters. Therefore, MinPts = 100 represents a representative choice balancing noise tolerance and structural integrity, and search distance = 70 km corresponds geographically to the typical spacing of oasis cities strung along piedmont corridors in Xinjiang.

From the results of OPTICS, the spatial structure (Figure 7) shows that there are two cores, four levels and several clusters. There are level 1 clusters (>3000 points), with the Northern Tianshan Mountains Urban Cluster and the Ili-Bortala Cluster comprising 59.1% of all clustered facilities. Altay and Kashgar, which serve as the main hubs in the tourism regions of the north and south, respectively, are at the level 2 clusters (900–1100 points). Transport corridors and intermediate oases are filled by level 3 clusters (300–600 points). At level 4, clusters (<200 points) are found in transitional/edge oasis areas. Noise points are mainly focused in the southern part of the Taklamakan Desert and in less populated counties, indicating that physical barriers still shape the accommodation system. Overall, clusters follow the demand gradient from gateway cities and core destinations toward secondary and tertiary areas.

3.3. Local Colocation Quotient (LCLQ) Analysis

3.3.1. Spatial Association Between Standard and Non-Standard Accommodation

The spatial relationship between standard and non-standard accommodation is dominated by the non-significant co-location pattern (Figure 8a). Across most counties and cities in Xinjiang, standard accommodation locations are not significantly associated with the prior distribution of non-standard accommodation, consistent with different site-selection patterns. In destinations with relatively limited tourism demand, this association is generally absent, whereas standard-hotel locations are more strongly associated with transport accessibility, business demand, and city size.

Significant attraction effects are confined to a few destinations with distinctive cultural identity and clear tourism orientation, including Kashgar, Yining, Tashkurgan, Xinyuan, and Tekes, with Kashgar containing the largest cluster of significant co-location points. This spatial selectivity carries important theoretical implications: the positive spatial association between prior non-standard accommodation agglomeration and subsequent standard hotel investment manifests only at nodes where cultural IP is sufficiently strong, and the tourism market has achieved a certain scale, rather than being pervasive across the entire region. This is consistent with the “market threshold” concept in central place theory—standard hotel investment requires a minimum demand scale, and only when non-standard accommodation has demonstrated destination market viability do conditions for standard hotel entry materialize.

In Kashgar Old City, spatial association exists between culturally embedded homestay and guesthouse clusters surrounding traditional Uyghur neighborhoods and subsequent standard hotels seeking concentrated visitor flows. Similar patterns appear in ethnic-cultural destinations, border tourism areas, and grassland tourism zones. Within Urumqi, attraction effects concentrate in the convention-exhibition district and high-speed rail area, where non-standard accommodation typically precedes standard hotels. Significant attraction points are frequently surrounded by non-significant co-location points, indicating that the spatial association effect of non-standard accommodation attenuates with distance from the urban core.

The reverse relationship presents a more balanced pattern (Figure 8b). In regional center cities such as Urumqi and Korla, spatial association exists between prior standard accommodation agglomeration and subsequent non-standard accommodation emergence, potentially related to standard hotels providing stable visitor flows and mature service infrastructure. However, this association remains spatially limited. In culturally resource-rich destinations such as Kashgar Old City, non-standard accommodation continues to follow the spatial distribution of cultural landscapes rather than existing hotel clusters. Overall, standard accommodation is more sensitive to market-size signals and urban service systems, while non-standard accommodation is more strongly associated with distinctive cultural environments and community-based tourism resources.

3.3.2. Hierarchical Differentiation Within Standard Accommodation

The spatial relationship between mass-market standard and premium standard accommodation is likewise dominated by non-significant co-location, though attraction effects emerge in several urban cores (Figure 8c). Significant hotspots concentrate in the central urban areas of Urumqi, Hotan, Korla, and Kuytun, where spatial association exists between prior premium-hotel distribution and subsequent mass-market standard-accommodation growth. This pattern is consistent with these locations sharing high destination quality, purchasing power, and business demand; it does not establish that premium hotels influence subsequent market entrants. The association primarily appears in regional-center cities where accommodation development is closely linked to commercial and administrative functions.

By contrast, major scenic-area destinations typically exhibit non-significant co-location. Constrained by land limitations and ecological protection requirements, premium hotels near natural scenic areas remain relatively dispersed, rarely achieving densities sufficient to produce measurable association effects. The contrast between urban centers and scenic-area destinations highlights different organizational logics: urban accommodation systems tend to reflect hierarchical market structures, while accommodation patterns near natural scenic areas are more strongly constrained by tourism resource distribution.

The reverse relationship shows weaker spatial association (Figure 8d). Only a limited number of counties contain significant co-location points, indicating limited spatial association between prior mass-market standard-accommodation distribution and subsequent premium-hotel locations. Synthesizing Figure 8c,d, the standard-accommodation sector exhibits an asymmetric pattern: the premium-to-mass-market association is stronger than the reverse association. This asymmetry describes the observed spatial pattern and should not be interpreted as an information flow or a location-decision mechanism.

3.3.3. Hierarchical Differentiation Within Non-Standard Accommodation

The relationship between mass-market non-standard and premium non-standard accommodation exhibits the strongest spatial independence among all accommodation categories (Figure 8e). In virtually all counties in Xinjiang, prior premium non-standard accommodation distribution produces almost no measurable association with mass-market non-standard location choices. This pattern relates to the limited numbers and dispersed distribution of premium homestays and boutique accommodation, which rarely form clusters sufficient to produce statistically identifiable association effects.

Significant association effects appear only in a few tourism destinations, including Burqin, Xinyuan, Habahe, Tekes, and Yining. These destinations share the common characteristics of strong tourism branding and high-quality natural landscapes. In Burqin, premium homestay and boutique accommodation clusters around Kanas and Hemu have formed identifiable tourism brands, with spatial association to subsequent mass-market non-standard accommodation growth. Results indicate that premium non-standard accommodation produces measurable spatial association effects only when supported by tourism resources of sufficient quality and scale.

Figure 8f presents the contrasting pattern and is the only relationship category dominated by association effects. Significant associations between mass-market and premium non-standard accommodation concentrate in the same destinations as Figure 8e. This pattern does not establish spatial guidance from mass-market to premium accommodation. The overlap is consistent with both categories being associated with established tourism demand, community participation, visitor-reception capacity, high-quality natural resources, and mature tourism markets, rather than with direct inter-hierarchical interaction.

3.3.4. Comprehensive Hierarchical Differentiation

The relationship between mass-market and premium accommodation is dominated by non-significant co-location (Figure 8g). Significant association effects are confined to a few destinations, most typically Burqin and Urumqi’s central urban area. In Burqin, premium hotel and boutique accommodation clusters around Kanas and Hemu coexist with subsequent mass-market accommodation growth; in Urumqi, association effects concentrate in the urban core where tourism demand and consumption capacity are highest. Beyond these, most counties show virtually no evidence of strong association between premium accommodation and mass-market location choices.

The reverse relationship presents an apparently balanced distribution (Figure 8h), yet significant co-location points are extremely sparse, appearing only in a few locations such as Habahe, Kashgar, Tashkurgan, and Urumqi’s Shayibake District. Most classifications are driven by non-significant co-location points, indicating that mass-market accommodation’s prior distribution provides extremely limited reference for premium accommodation development. Premium accommodation maintains high selectivity, more strongly associated with destination quality, tourism resources, market structure, and brand value than with the spatial density of lower-tier facilities.

Synthesizing Figure 8g,h, hierarchical differentiation in Xinjiang’s accommodation industry is primarily shaped by resource endowment and market-condition differences rather than direct interaction between accommodation categories. Premium accommodation, constrained by higher investment thresholds and stricter location requirements, exhibits highly selective spatial distribution. Mass-market accommodation operates under broader market conditions and thus displays wider spatial coverage. Where the two categories coexist, their spatial overlap is most pronounced in destinations with exceptional tourism resources or mature urban markets, suggesting that both categories respond to the same underlying locational advantages rather than to each other’s spatial distribution.

3.4. Feature Distribution and Preliminary Comparison of Categorized Accommodation Types

To compare site-selection preferences for each grade and type, we examined the distributions of 23 explanatory factors for nine accommodation types: mass-market standard (MS), mass-market non-standard (MNS), mass-market total (MM), premium standard (PS), premium non-standard (PNS), premium total (PM), standard total (ST), non-standard total (NS), and all accommodation. The results are summarized below, by six thematic dimensions.

3.4.1. Elevation Gradient and Ecological Foundation

The elevation violin plot (Figure 9a) shows a site-selection gradient from foothill plains to remote mountain areas. MS (mean = 900.53 m) and PS (966.45 m) cluster at lower elevations, MNS (1047.91 m) and NS (1054.51 m) occupy intermediate elevations, and PNS (1302.84 m) occurs at the highest elevations. Slope increases from 0.90° for MS to 3.50° for PNS. Vegetation cover rises from 0.248 for MS to 0.538 for PNS. Temperature declines from 10.63 °C for MS to 6.10 °C for PNS, while precipitation increases from 133.34 mm to 243.98 mm. PS clusters closely with MS in elevation, slope and precipitation, reflecting the predominantly urban location of both categories. PNS is absent from Urumqi and is concentrated in Ili (50.0%) and Altay (33.9%).

3.4.2. Urban-Rural Spatial Binary Polarization

Economic factors show urban-rural polarization. Urbanization intensity for ST (36.77) is approximately twice that of NS (18.35). Regional GDP shows a similar contrast (415.22 billion versus 205.51 billion; ratio = 2.02), and the nighttime light index is 3.23 times higher for ST than for NS (16.67 versus 5.16). PS and MS are almost identical economically, suggesting that PS achieves premium positioning through product quality within the same urban niche (Figure 9g–i,s).

3.4.3. Common Foundation of Transportation Networks and Selective Node Dependence

Distance to main roads varies only slightly, approximately 0.10–0.61 km, suggesting that road accessibility functions as a common threshold. Standard accommodation lies closer to railway stations, with a mean distance of approximately 23 km, than non-standard accommodation, which averages approximately 30 km. MNS is closest to parking lots (0.11 km), reflecting the self-drive tourism market. PS is closest to airports (26.59 km), reflecting strategic positioning for business travelers (Figure 9k–o).

3.4.4. Three-Pole Pattern of Tourism Resource Response

Tourism resources show three response modes. PS is closest to high-level (9.52 km) and low-level (5.44 km) scenic areas, indicating scenic brand attachment. MNS, NS and PNS are more closely coupled with rural tourism sites, with the median distance from MNS/NS to such sites reaching 2.50 km. MS/ST occupy intermediate positions that balance scenic access with urban services (Figure 9u–w).

3.4.5. Extreme Polarization of Premium Accommodation

PS exhibits urban-scenic dual attachment, whereas PNS shows ecological embeddedness, with the highest elevation (1302.84 m), steepest slope (3.50°) and lowest urbanization intensity (12.48). PS achieves premium positioning through proximity to both urban and scenic resources. PNS relies on deeper ecological and cultural immersion, indicating a different premium-location strategy (Figure 9a–c,g).

3.5. XGBoost-SHAP Analysis of Nonlinear Driving Mechanisms

3.5.1. Model Validation and Predictive Performance

All 18 models performed well under the random-split evaluation (Figure 10). Under MHWD, test-set R² ranged from 0.912 to 0.990, with MHWD-Total performing best (R² = 0.9903, RMSE = 24.72). MHWD-PNS and MHWD-PM performed less strongly (R² = 0.9121 and 0.9342), probably due to the small PNS sample and high spatial dispersion. Under SSI, test-set R² ranged from 0.9127 to 0.9908 and was comparable with MHWD, while RMSE and MAE were lower. This difference suggests that SSI captures a more stable geographic pattern, whereas MHWD incorporates review-based platform demand variation.

To characterize residual spatial structure, we performed a Global Moran’s I diagnostic for all 18 models. Low residual autocorrelation indicates that little spatial structure remains in the fitted residuals. However, this post-fit diagnostic cannot demonstrate the absence of spatial information leakage or replace spatially blocked cross-validation. Because the retained KDE models were evaluated with a random split, their R² values should be interpreted as in-domain fit rather than as estimates of geographic transferability [55].

The diagnostics reveal two key points. First, the dependent variable exhibits significant positive spatial autocorrelation, confirming the inherently agglomerated nature of accommodation density (

Table 9. Residual spatial autocorrelation diagnostics for 18 XGBoost models.

Category	Moran’s I (Y)	p	Moran’s I (Resid.)	p	Reduction
MS-MHWD	0.977	<0.001	−0.021	<0.001	97.80%
MS-SSI	0.945	<0.001	0.003	0.285	99.70%
MNS-MHWD	0.988	<0.001	−0.017	0.006	98.30%
MNS-SSI	0.979	<0.001	−0.023	<0.001	97.70%
MM-MHWD	0.986	<0.001	0.044	<0.001	95.60%
MM-SSI	0.978	<0.001	0.079	<0.001	91.90%
PS-MHWD	0.904	<0.001	−0.100	<0.001	89.00%
PS-SSI	0.839	<0.001	−0.090	<0.001	89.20%
PNS-MHWD	0.743	<0.001	−0.124	0.008	83.30%
PNS-SSI	0.805	<0.001	−0.124	<0.001	84.60%
PM-MHWD	0.918	<0.001	−0.097	<0.001	89.40%
PM-SSI	0.836	<0.001	−0.089	<0.001	89.30%
ST-MHWD	0.983	<0.001	−0.009	0.062	99.10%
ST-SSI	0.953	<0.001	0.013	0.014	98.60%
NS-MHWD	0.989	<0.001	−0.014	0.014	98.60%
NS-SSI	0.98	<0.001	−0.016	0.006	98.40%
ALL-MHWD	0.989	<0.001	0.051	<0.001	94.90%
ALL-SSI	0.98	<0.001	0.085	<0.001	91.30%

). The MHWD series consistently exceeds the SSI series because review weighting amplifies agglomeration signals, thereby highlighting density contrasts.

Second, residual spatial autocorrelation attenuates by 83.3–99.7% to a range of −0.124 to +0.085, approaching spatial randomness. Twelve models show negative Moran’s I values, indicating mild “checkerboard” dispersion rather than unmodeled clustering. Given that |I| < 0.1 constitutes “negligible spatial autocorrelation”, the MS-SSI model yields the optimal result with a residual Moran’s I of +0.003 (p = 0.285), indistinguishable from a random distribution.

In conclusion, the residual diagnostics indicate that the fitted models leave relatively limited spatial autocorrelation. The feature-importance rankings and interaction patterns should nevertheless be interpreted as in-domain, exploratory associations, because residual Moran’s I does not substitute for spatially blocked validation or establish geographic transferability.

3.5.2. Hierarchical Structure of Driving Factors and Dual-Perspective Differentiation

SHAP global feature importance revealed significant hierarchical differentiation among the 23 driving factors, with fundamental divergence between the demand-side (MHWD) and supply-side (SSI) perspectives (Figure 11). Feature importance was calculated as Mean |SHAP|; relative percentages are comparable only within each model.

Under MHWD, the Proportion of Tertiary Industry (PTI) occupied a dominant position across all standard accommodation types (24–29%), far exceeding the second-ranked factor. This dominance reflects the deep economic-structure logic underlying accommodation distribution: regions with high PTI possess developed tourism industry chains, adequate labor supply, and mature consumption markets that collectively constitute prerequisites for sustained market heat. However, PTI’s importance collapsed in non-standard types—it did not enter the top six in MHWD-MNS and fell below 3% in MHWD-PNS—demonstrating that non-standard accommodation location logic has transcended traditional economic-structure constraints.

Complementary to PTI, Urbanization Intensity (UI) and Distance to Railway Station (D_Rail) formed a triadic driving structure (PTI–UI–D_Rail) for standard accommodation: PTI provides the industrial foundation, UI represents market capacity, and D_Rail determines medium-to-long-distance passenger-flow import efficiency. This structure found its most typical confirmation in the Ürümqi–Shihezi–Kuitun Northern Xinjiang railway corridor. In contrast, non-standard accommodation under MHWD exhibited a three-dimensional framework of administrative-center anchoring, tourism-resource coupling, and transportation-node support: D_RTDS ranked first in MHWD-MNS (18.79%), D_County second (14.47%), and D_Park third (13.82%); in MHWD-ST, D_County rose to first (22.02%) and D_Air entered the top three (11.17%). Premium non-standard accommodation (MHWD-PNS) further highlighted ecological embeddedness: D_RTDS (13.85%) and D_Riv (13.76%) jointly occupied the top two, with all economic factors failing to enter the top six.

Under SSI, Population Density (PD) replaced PTI as the primary driver for most types (16–19%), reflecting investors’ pragmatic consideration of market scale. The Density of Catering and Shopping POI (CSP_Den) performed prominently in non-standard types, ranking first in both SSI-MNS (18.74%) and SSI-ST (24.36%), revealing supply-side sensitivity to lifestyle-service amenities.

A comprehensive comparison yielded three differentiation patterns: (1) economic-structure-driven (MHWD) versus population-market-driven (SSI); (2) driving consistency of standard accommodation versus driving heterogeneity of non-standard accommodation across perspectives; (3) ecologically dominant driving for premium non-standard accommodation remaining stable across both perspectives (D_RTDS/D_Riv in MHWD-PNS; D_Rail/D_Riv/Elev in SSI-PNS), providing solid empirical support for ecological conservation-priority policies.

3.5.3. Nonlinear Marginal Effects and Threshold Characteristics

SHAP partial dependence plots revealed that driving factors exhibited complex threshold mutations, interval reversals, and saturation effects rather than linear marginal contributions (Figure 12 and Figure 13). Thresholds were identified at feature values where Lowess-fitted partial dependence curves crossed zero.

Under MHWD, PTI exhibited the most significant threshold: below approximately 67%, its marginal contribution was negative; above 67%, it turned sharply positive. This corresponds to the structural inflection point where regional economies transition from industry-dominated to service-dominated. The threshold for premium standard accommodation was slightly higher (~70%), indicating stricter service-maturity requirements. UI thresholds displayed gradient differentiation: 37.9 (MHWD-MS), 39.8 (MHWD-ST), and 45.1 (MHWD-PS), with increasing thresholds reflecting higher urbanization requirements for premium hotel operations.

D_Rail revealed a three-stage structure in MHWD-MS: positive within 0–5.8 km (core station radiation radius), negative in 5.8–126 km, and slight easing beyond 126 km (approximate inter-station spacing on Xinjiang’s railway network). This dual-threshold structure was stable across standard types (5.3–5.8 km/123–126 km) but weakened in premium standard accommodation (single threshold ~5.1 km), reflecting premium clientele’s preference for air travel.

In non-standard accommodation, D_RTDS exhibited a single threshold at 3.0 km (MHWD-MNS)—the maximum effective radiation radius of rural tourism demonstration sites. D_County displayed a negative–positive–negative wave structure: negative below 0.56 km (direct competition with urban hotels), positive in 0.56–51.63 km (county-town support with rural independence), and negative beyond 51.63 km (exceeding effective service radius). Elevation showed a single threshold at 1360 m (MHWD-ST), matching the elevation range of major valley-grassland scenic areas (Ili Valley: 1200–1500 m; Kanas: 1300–1400 m) and representing the upper limit of developable elevation.

Under SSI, CSP_Den in non-standard accommodation exhibited an inverted-U structure with dual thresholds (1224.5 and 53,590.6 units/km² in SSI-MNS): moderate commercial density facilitates operations, but extreme density signals full urbanization where non-standard accommodation loses its rural scarcity premium. PD exhibited monotonic increasing patterns with type-differentiated thresholds: 3.37 persons/grid (SSI-MS), 2.29 (SSI-MM), and 1.95 (SSI-ST), confirming that non-standard accommodation survives on niche markets requiring lower population-scale support.

3.5.4. Interaction Effects and Synergistic Driving Mechanisms

Based on TreeSHAP pairwise interaction decomposition, this section examined non-additive effects of dual-factor interactions on accommodation density (Figure 14 and Figure 15). The top 15% of 253 potential pairs were screened for focused analysis.

Under MHWD, interaction effects constituted substantial proportions of total feature contributions: in MHWD-ALL, PTI’s interaction component reached 37.5%, while FVC and MAT exhibited interaction-dominant structures (60.5% and 61.9%, respectively). This pattern is consistent with the interpretation that natural environmental factors do not operate in isolation but are modulated by economic and accessibility conditions.

The strongest MHWD interactions fell into two groups. For standard accommodation, PTI × D_County was the dominant pair (18.74 in MHWD-ALL; 6.94 in MHWD-PS), indicating that the association between tertiary industry share and accommodation density is substantially amplified near county seats. For non-standard accommodation, D_RTDS × D_Park ranked first in MHWD-MNS (3.06), and D_County × Elev was the strongest pair in MHWD-ST (3.68). The latter suggests that the joint configuration of administrative proximity and elevation—rather than either factor alone—may be particularly relevant for non-standard accommodation density, though the underlying mechanism warrants further investigation with longitudinal data.

In premium non-standard accommodation (MHWD-PNS), the interaction structure shifted: D_Riv × D_LLSA became the strongest pair (0.49), followed by D_RTDS × FVC (0.21). This is consistent with a compound resource configuration in which water-body proximity and peripheral recreation amenities jointly condition premium homestay density. Notably, antagonistic (negative) interactions also appeared: Elev × D_Rail was negative in the high-elevation, far-from-railway interval in MHWD-PNS, which may indicate that in highly remote ecological settings, railway absence is not associated with reduced density—though this pattern requires cautious interpretation given the small sample (N = 186).

Under SSI, PD × D_Rail was the strongest interaction for standard types (0.028–0.031), suggesting that the association between population density and supply-side density is stronger near railway stations. For non-standard types, CSP_Den × Elev dominated (0.079 in SSI-ST), indicating that the joint configuration of commercial amenity density and elevation is particularly relevant for non-standard accommodation supply.

Three interaction typologies were identified: (1) Amplifying—both factors have positive main effects and their joint value exceeds the additive sum by >50% (e.g., PTI × D_County, PD × D_Rail); (2) Compensating—a weak main-effect factor gains effective influence through interaction with a strong factor (e.g., FVC × PTI, Elev × CSP_Den); (3) Threshold—both factors possess individual zero-crossing thresholds, and the joint interaction turns positive only when both simultaneously exceed their critical values (e.g., D_RTDS × D_Park). These typologies provide a descriptive taxonomy for the observed non-additive patterns; causal interpretation of the underlying mechanisms requires further evidence beyond cross-sectional SHAP decomposition.

4. Discussion

4.1. Two Locational Regimes: Spatial Differentiation of Standard and Non-Standard Accommodation

The most central finding of this study is that standard and non-standard accommodation follow systematically different locational logics, a differentiation consistently observed across the full nine-category × two-perspective matrix. For standard accommodation (MS, PS, ST), SHAP importance is dominated by PD (16–18%), UI (14%), and D_Rail (7–11%)—population density, urbanization intensity, and railway accessibility—closely resembling the “urban centrality + transport node” gravitational field predicted by central place theory and location theory [8,10]. For non-standard accommodation (MNS, PNS, NS), the importance spectrum shifts toward D_RTDS (distance to rural tourism demonstration sites), D_Riv (distance to water bodies), Elev (elevation), and D_LLSA (distance to low-grade scenic areas), with economic variables contributing minimally. PNS-SSI is most extreme: D_Rail (13.8%) > D_Riv (11.0%) > Elev (9.4%), with economic variables barely entering the top six, exhibiting deep ecological embeddedness.

This differentiation is highly consistent with agglomeration economics predictions [8]. Standard hotels primarily exploit urbanization economies—diversified labor pools, shared supplier networks, business visitor flows—and therefore anchor in city centers. Non-standard accommodation follows resource-oriented logics, where natural and cultural assets compensate for thinner local markets [16]. TALC further predicts that as destinations evolve from exploration toward consolidation, the accommodation mix shifts from resource-dependent to market-oriented [16,17]. Xinjiang is currently in the early-to-middle stages of this transition: the explosive growth of non-standard accommodation in the Ili Valley and Altay (since 2019) corresponds to TALC’s “involvement → development” stage, while Ürümqi’s standard hotel system has entered the “consolidation” stage.

The contrast between Burqin County and Ürümqi is the most vivid illustration of this differentiation. Ürümqi has 215 premium standard establishments (including 29 five-star) but zero premium non-standard establishments; Burqin has 49 premium non-standard establishments (first among all counties in Xinjiang) but very few premium standard establishments. The two represent entirely different visitor structures and market logics: Ürümqi is dominated by business travelers, where standardized service and brand reputation are core competitiveness; Burqin is dominated by sightseeing and leisure travelers, where local culture and landscape ambiance are core attractions.

4.2. Theoretical Interpretation of Nonlinear Thresholds

The threshold patterns observed in the SHAP partial-dependence plots represent exploratory spatial-economic associations with possible theoretical correspondences. Given the cross-sectional and exploratory nature of the analysis, they should not be interpreted as causal effects or as definitive evidence of the underlying mechanisms. This study identifies three categories of thresholds and their theoretical interpretations:

Market-size thresholds. PTI (tertiary industry share) produces a positive-to-negative transition at approximately 67% (MS-MHWD), corresponding to the qualitative transformation point where economic structure shifts from industry-dominated to service-dominated. Below this threshold, cities may lack the service-sector ecology to sustain chain hotel operations; above it, business visitor flows and service-sector labor pools plausibly render hotel investment viable. This pattern is consistent with Christaller’s market threshold concept [8], which posits that facility existence requires a minimum demand scale. Similarly, PD thresholds at 3.37 persons/grid (MS-SSI) and 2.28 persons/grid (ALL-SSI) are consistent with the notion of a minimum population base required for accommodation facility operations.

Ecological constraint thresholds. The elevation threshold of 1360 m (ST-MHWD) broadly corresponds to the ecological habitability boundary of the Ili Valley and Kanas Basin—above this elevation, winter heating costs, construction difficulty, and visitor accessibility may deteriorate sharply, and the economic viability of accommodation investment is likely to decline substantially. The threshold effect of D_Riv is consistent with a strong constraint of water bodies on resort-type development in arid regions: in environments with annual precipitation of only 133–244 mm, distance to rivers/lakes may be a key condition governing landscape value and water-supply feasibility.

Agglomeration shadow effects. D_Rail produces dual thresholds at 5.8 km (MS-MHWD) and 90.6 km (ALL-SSI). Within 5.8 km, the effect is positive (passenger-flow dividends from railway stations); between 5.8 and 90.6 km, the effect is negative (the “intermediate zone” that is neither within railway-station radiation range nor in the urban core); beyond 90.6 km, the effect attenuates (remote tourism destinations such as Kanas and Tashkurgan do not depend on railways). This “agglomeration shadow” pattern has a theoretical correspondence in evolutionary economic geography [19,29]: early agglomeration around transport nodes is thought to create spatial lock-in, and entrants deviating from the locked-in zone may face increasing costs.

4.3. Directional Spatial Associations and Literature Comparison

The directional spatial associations revealed by LCLQ analysis exhibit hierarchical characteristics consistent with central place theory predictions [8]. The standard → non-standard directional association concentrates in culturally distinctive destinations such as Kashgar Old City, Yining, Tashkurgan, Xinyuan, and Tekes, indicating that in mature tourism areas, significant positive spatial association exists between prior non-standard accommodation agglomeration and subsequent standard hotel investment. This pattern is consistent with the hierarchical diffusion pattern found by Chen & Wei (2025) [42] in Chinese urban hotels, but this study further reveals the spatial selectivity of this association in arid regions—it manifests only at nodes where cultural IP is sufficiently strong, rather than being pervasive.

The reverse (standard → non-standard) association concentrates mainly in regional center cities such as Ürümqi and Korla, indicating that in administrative centers, spatial association exists between standard hotel agglomeration and subsequent non-standard accommodation emergence, but in cultural core areas, non-standard accommodation continues to follow cultural landscapes rather than hotel clusters.

The premium → mass-market directional association appears in the urban cores of Ürümqi, Hotan, Korla, and Kuytun, and is consistent with shared destination quality, market structure, and demand conditions [65]. However, the mass-market → premium direction appears only sporadically, indicating that premium-accommodation locations are more strongly associated with destination quality and market structure than with the density of lower-tier properties.

It must be reiterated that all “associations” described above are spatial-statistical descriptions and do not constitute causal inference. Temporal precedence is a necessary but insufficient condition; unobserved third-party factors (e.g., government tourism functional-zone designation, scenic-area rating upgrades, major transport infrastructure openings) may simultaneously drive location choices across multiple categories.

4.4. Sustainability Implications and Category-Differentiated Governance

The findings of this study carry direct policy implications for sustainable tourism development in arid regions. The identified nonlinear thresholds provide spatially explicit reference points for planning:

Ecological feasibility boundaries. The 1360 m elevation threshold and D_Riv threshold effects jointly define the ecological feasibility envelope for accommodation development. Under the SDG 6 (Clean Water) and SDG 15 (Life on Land) frameworks, development beyond these boundaries faces dual risks of water-resource unsustainability and ecological degradation. Planning tools can incorporate these thresholds as “red-line” parameters to monitor whether accommodation growth is encroaching on ecologically sensitive zones.

Category-differentiated governance. The locational logic of standard accommodation (urban centrality + transport nodes) implies that its expansion is primarily market-driven, and policy intervention should focus on intra-urban land-use coordination and transport connectivity. The locational logic of non-standard accommodation (tourism resources + ecological assets) implies that its expansion directly touches ecologically sensitive zones, and policy intervention should focus on light-touch registration, ecological monitoring, and water-resource auditing. The deep ecological embeddedness of premium non-standard accommodation (PNS)—highest elevation, steepest slopes, lowest urbanization intensity—makes it the highest ecological-risk category, requiring ecological certification, carrying-capacity limits, and digital visibility support.

Cluster-hierarchy planning. The four-tier cluster structure identified by OPTICS provides a spatial framework for differentiated investment: Level 1 (>3000 facilities) requires quality upgrading and capacity management; Level 2 (900–1100) requires connectivity and joint marketing; Level 3 (300–600) requires infrastructure and incentives; Level 4 (<200) requires targeted incubation or development restraint.

4.5. Limitations and Future Directions

This study has several limitations. First, data originate from the Ctrip platform and may not fully cover unregistered or non-platformized accommodation [27,66,67]; diamond ratings and review counts, as platform proxy indicators, may deviate from true market demand. Second, XGBoost-SHAP identifies statistical association rather than causation [68,69,70], and some interaction effects may be influenced by confounding variables; the temporal precedence setting in LCLQ is a necessary but insufficient condition for causal inference. Third, the cross-sectional design limits characterization of dynamic evolutionary processes in the accommodation system. Future research could incorporate multi-period panel data to track temporal evolution of inter-category spatial associations, employ quasi-experimental designs or instrumental variable approaches to explore causal pathways, and conduct comparative studies with other western arid regions (e.g., the Hexi Corridor, Qaidam Basin) to test the external applicability of the present findings.

5. Conclusions

Drawing on 12,073 accommodation establishments from the Ctrip platform, this study constructed a staged analytical framework of “description → decomposition → explanation” and systematically examined the spatial distribution heterogeneity, directional spatial associations, and nonlinear driving mechanisms of accommodation facilities in Xinjiang across the full matrix of nine categories × two perspectives. The principal conclusions are as follows:

(1) Accommodation facilities in Xinjiang exhibit a highly concentrated “single-core, multi-center” structure, with Ürümqi as the absolute core and Kashgar and Yining as secondary anchors, expanding rapidly toward the Ili Valley and Altay since 2019. OPTICS clustering identified 13 spatial urban groups, and parameter sensitivity analysis confirmed that this structure is highly robust within the range of search distances 70–100 km and MinPts 100–150 (ARI ≥ 0.98).

(2) Standard and non-standard accommodation follow systematically different locational logics. The SHAP importance of standard accommodation is dominated by population density, urbanization intensity, and railway accessibility, consistent with the urban-centrality gravitational field predicted by central place theory and location theory; non-standard accommodation shifts toward ecological resource variables including distance to rural tourism demonstration sites, water-body distance, and elevation, with economic variables contributing minimally. This differentiation is consistently observed across all 18 models in the nine-category × two-perspective matrix.

(3) Driving factors exhibit pronounced nonlinear associations: the proportion of tertiary industry shows a positive-to-negative transition at approximately 67%; elevation is associated with an inflection near 1360 m for non-standard accommodation; and railway-station distance and distance to rural tourism demonstration sites show nonlinear association patterns. Residual Moran’s I diagnostics (reduction 83–99.7%) suggest that these findings are unlikely to be artifacts of spatial autocorrelation, while we acknowledge that this diagnostic complements rather than replaces spatially blocked validation.

(4) Directional local co-location quotient analysis reveals hierarchical spatial associations among categories: the positive association between prior non-standard accommodation agglomeration and subsequent standard hotel investment appears only at nodes with sufficiently strong cultural IP (Kashgar Old City, Yining, Tashkurgan) rather than being pervasive; hierarchical association within standard accommodation is characterized by unidirectional spatial information flow; and spatial association of premium non-standard accommodation is constrained by extremely high resource thresholds, manifesting only in a few destinations such as Burqin and Xinyuan.

The contribution of this study lies in integrating XGBoost–SHAP, KDE, OPTICS, and LCLQ into a staged analytical framework and systematically implementing it across the full nine-category × two-perspective matrix of arid-region accommodation systems, thereby identifying two locational regimes, nonlinear thresholds, and directional spatial associations that provide a spatially explicit evidence base for category-differentiated sustainable tourism governance. In practical terms, these findings may offer quantitative references for ecological feasibility screening, differentiated management pathway design, and investment prioritization discussions for accommodation development in arid regions, though their policy translation requires longitudinal monitoring data and local governance context.