Analysis of Hotel Reviews and Ratings with Geographical Factors in Seoul: A Quantitative Approach to Understanding Tourist Satisfaction

Abstract

This study examines how hotel characteristics and urban spatial context influence tourist satisfaction in Seoul, South Korea, by integrating sentiment analysis of online reviews with regression modeling. Drawing on 4500 TripAdvisor reviews from 75 hotels, sentiment scores were extracted using aspect-based sentiment analysis, and two regression approaches, ordinary least squares (OLS) and spatial autoregressive combined models, were applied to evaluate how hotel specific features, such as the age and scale of the hotels and room rates, and their geographic characteristics, such as the proximity to airports and cultural landmarks, affect both emotional sentiment and formal hotel ratings. The OLS model for sentiment scores identified the scale and rating of the hotels as well as the proximity to the airports as key predictors. Additionally, the spatial autoregressive combined model was also statistically significant, suggesting spatial spillover effects. A separate model for the traditional rating revealed weaker associations, with only the hotel’s opening year reaching significance. These findings highlight a divergence between emotional responses and structured ratings, with sentiment scores more sensitive to spatial context. This study offers practical implications for hotel managers and urban planners, emphasizing the value of incorporating spatial factors into hospitality research to better understand the tourist experience.

1. Introduction

The rapidly developing hospitality industry of South Korea, coupled with the widespread use of digital platforms, has dramatically changed how tourists select and evaluate hotels. Traditional rating systems offer standardized summaries of hotel quality but often fail to reflect the full spectrum of guest experiences. In contrast, online text reviews capture emotional nuances that numeric ratings may overlook. Sentiment analysis techniques allow for the extraction of these subjective insights from text reviews, and when merged with geographic data, give an avenue for examining how geographic context influences guest perception.

Despite the growing interest in sentiment analysis within tourism research, the spatial dimension of tourist sentiment is relatively under-explored. This study addresses that gap with an empirical, quantitative approach that analyzes the relationship between hotel features, spatial proximity, and guest ratings in Seoul. It integrates sentiment analysis of online ratings for hotels with spatial regression modeling to investigate how internal features (e.g., the number of guest rooms, costs) and geographic characteristics such as proximity to transport infrastructure or tourist attractions affect emotional evaluations of hotel stays.

The analysis considers three main categories of variables: (1) hotel-specific characteristics such as opening year, size, star rating, and price; (2) geographic features including accessibility to public transportation and distance to key landmarks like airports and the palace; and (3) guest sentiment, quantified through text-based review scores. As Wu et al. [1] suggested, guest emotions are shaped by subtle spatial stimuli such as neighborhood atmosphere and local ambiance, which are not captured by traditional ratings alone. Sentiment scores are less rigid and more subjective than formal ratings and release patterns of satisfaction that might otherwise be overlooked.

While other previous studies have explored sentiment analysis or urban accessibility individually, few have integrated these elements within a single analytical framework. As Latinopoulos [2] notes, ignoring spatial context can lead to incomplete tourist satisfaction models. This study builds on that insight by incorporating spatial lag models to account for spatial dependencies in guest reviews. Although regression analysis is widely used in tourism studies, few models explicitly consider spatial heterogeneity.

To this end, this study applies both ordinary least squares (OLS) and spatial autoregressive combined models to assess the combined effect of hotel internal characteristics and urban space features on guest reviews. By comparing sentiment-derived scores with formal hotel ratings, the study illustrates the diverse ways in which tourists respond to service quality and spatial environments.

2. Literature Review

Over recent years, numerous studies have explored the relationship between customer sentiment and hotel ratings. The foundational framework for understanding satisfaction is Expectation–Disconfirmation Theory (EDT), originally proposed by Oliver [3], which posits that satisfaction results from the gap between prior expectations and actual performance. EDT has been widely applied in the hospitality industry for understanding how guests evaluate service quality, price equity, and overall hotel experiences (Oh [4], Baker & Crompton [5]). A panel dataset of more than 44,000 guests demonstrated that expectancy–disconfirmation influences predominant placebo influences and that prices beyond expectations result in reduced ratings in reviews even under constant service levels (Abrate, G., Quinton, S., & Pera, R. [6]). In experimental settings, positive disconfirmation, where actual experience surpasses expectations, was found to increase guests’ willingness to post favorable reviews, indicating behavioral reinforcement via review ratings (Li, Meng & Hudson [7]).

Sentiment, as it occurs in online reviews, instantiates these disconfirmations as text. Geetha et al. [8] confirmed that sentiment polarity, positive or negative, accurately predicts hotel star ratings. Lai et al. [9] identified an asymmetric effect, showing that negative sentiment has a larger effect than positive feedback, consistent with the broader negativity bias in consumer psychology. Al-Natour and Turetken [10] claimed that sentiment scores extracted from guest reviews can at times represent customer experiences better than conventional structured ratings. Yet, most of such studies use global OLS models and neglect spatial heterogeneity and geographic effects intrinsic to tourism.

The addition of spatial theory, in this case Tobler’s First Law of Geography “everything is related to everything else, but near things are more related than distant things” (Tobler [11]), provides guest satisfaction modeling with a significant dimension. Spatial proximity to transportation, landmarks, or other highly rated hotels could affect both guest opinion and satisfaction scores. For instance, Wang and Zhou [12] used spatial econometric models to demonstrate that spatial dependence of hotel satisfaction dominates in San Francisco, while Latinopoulos [2] determined that proximity to tourist and transit hubs constantly reinforces hotel ratings. These spatial elements affect expectation–disconfirmation processes by influencing experience formation and evaluation.

Parallel to this, sentiment analysis studies have emerged to uncover patterns of dis-satisfaction and satisfaction in comments. Approaches range from polarity-based classification to topic modeling to enable researchers to quantify satisfaction and identify key hotel attributes such as location, cleanliness, and service (Chen et al., [13]). However, as Lu and Stepchenkova [14] and Xiang et al. [15] emphasize in their meta-reviews, most sentiment-based research does not include a geographic outlook. Recent advances in geotagged review analytics offer an upcoming window to connect text and space. For instance, Kádár and Gede [16] geotagged photographed tourist mobility. This supports the argument that emotion is not merely textual but also spatially embedded, offering a richer spectacle to understand tourist experience.

The relationship between sentiment score and formal rating has also been researched. Carreon et al. [17] identified a moderate relationship between star ratings and sentiment through Spearman and Kendall coefficients. Chen et al. [13] integrated sentiment analysis and the Kano-IPA model to examine the impact of hotel attributes towards satisfaction, which revealed multidirectional influences untapped by ratings only. But the studies tend to make space peripheral, not central, when in fact tourism behavior is spatial.

Research that has been performed to identify the determinants of hotel satisfaction has also been on the rise. Gunasekar and Sudhakar [17] examined hotel attributes such as location, comfort, and service quality, which confirmed their effect on guest rating. Another research that was more spatially explicit by Valenzuela-Ortiz et al. [18], who used spatial quantile regression to show that hotels that are located in protected natural areas or human capital investment areas receive better ratings. In particular, they also remarked that rural or mountain hotels are accessible to tourists who seek outdoor or adventure activities, whereas city hotels are preferred by location and infrastructure, suggesting space and satisfaction have complex associations.

Despite the growing literature linking sentiment, satisfaction, and space, few have endeavored to integrate these concepts within an overarching analytical framework. Most existing studies rely on linear regression or exploratory mapping without accounting for spatial dependence and without conceptualizing affective feedback from a spatially conscious perspective. Spatial econometric models, and spatial lag and spatial error models in particular, explicitly conceptualize the influence of the neighboring units and are therefore well-suited to model geographically situated reviews.

This study contributes to the literature by integrating sentiment analysis with spatial data through two regression models: an Ordinary Least Squares (OLS) model as a baseline, and spatial autoregression model to test for spatial spillover effects. Drawing on Oliver’s [3] EDT and Tobler’s Law [11], we argue that guest satisfaction is shaped by both emotional content and geographic proximity. Rather than examining sentiment polarity or hotel attributes independently, this study quantifies the influence of both textual and spatial variables on satisfaction. Incorporating spatial closeness to transport, attractions, and tourist infrastructure provides a more complete understanding of what drives tourist experience. In this way, this study not only expands upon earlier hospitality studies but contributes to the emerging field of spatially enabled sentiment analytics in tourism geography.

3. Data and Methods

3.1. Study Area and Data Preprocessing

This study analyzes the hotel reviews for Seoul, South Korea, collected from TripAdvisor. The dataset comprises 4500 reviews across 75 hotels, representing a range of accommodation types from low-cost lodging to upscale establishments. Figure 1 shows the locations of the hotels covered in the dataset.

Along with the review data collected from TripAdvisor, the coordinates of bus stops, subway stations, and convenience stores were downloaded from the Seoul Open Data Plaza, an official platform offering a wide array of public datasets. Buffer zones of 600 m were generated around the hotels to compute the availability of proximate facilities (i.e., bus stops, subway stations, and convenience stores) using QGIS 3.36 3, an open-source Geographic Information System software.

By using buffer analysis and spatial joins, every hotel was associated with the number of local amenities within walking distance. The selection of a 600 m buffer was informed by walkability and city accessibility studies where 400 to 800 m is typically applied as the pedestrian comfort zone. A 600 m radius is roughly a 7 to 10 min walk and detects amenities most likely to be in walking distance from tourists.

Apart from these buffer-based figures, distances and time directly between the hotels and salient points were calculated using a QGIS software. These points were Incheon and Gimpo airports, representing regional transport nodes, and Gyeongbokgung Palace, representing a popular culture and history destination. As discussed previously, spatial accessibility is a critical factor in determining tourist satisfaction, so we included variables, such as nearness to transit, convenience stores, and landmarks in the models of sentiment and ratings. Figure 2 shows the locations of these facilities and landmarks.

While this study draws exclusively on TripAdvisor reviews for sentiment analysis, several strands of evidence support the validity and representativeness of this data source in the context of Seoul’s tourism landscape. Although language and platform biases are common concerns in user-generated content research, Hale [19] found that TripAdvisor star ratings across multiple language groups are highly correlated, suggesting that even English-language reviews align closely with broader tourist sentiment patterns. Additionally, Chua and Banerjee [20] demonstrated high inter- and intra-reviewer consistency on TripAdvisor, reinforcing the platform’s reliability for aggregating guest experience data. More recently, Han and Anderson [21] showed that TripAdvisor exhibits lower review nonresponse bias than comparable platforms such as Google and Booking.com, and while its text reviews tend to be slightly more positive than structured survey responses, this bias is comparatively mild. Crucially, a study by Hong [22] on TripAdvisor reviews of Gyeongbokgung Palace in Seoul found that international tourists frequently voice concerns related to crowding, authenticity, accessibility, and site engagement, underscoring that free-text reviews reflect meaningful spatial experiences across diverse user groups. Taken together, these studies provide strong justification for using TripAdvisor sentiment data in urban hospitality analysis.

Table 1. List of variables (n = 75).

Variable Name	Description	$\mathit{\mu }$	$\mathit{\sigma }$
rating	Hotel rating	4.40	0.36
open_yr	Year open	2008.73	14.5499
n_rooms	Number of rooms in the hotel	214.08	139.32
dist_palace	Distance to Gyeongbokgung Palace	6690.16	4556.30
dist_icn	Distance to the Incheon Airport	62,619.63	4512.03
dist_gmp	Distance to the Gimpo Airport	27,545.75	5104.81
bus	Number of bus stops within 600 m	21.21	8.64
subway	Number of subway stations within 600 m	2.84	2.21
cvs	No. convenience stores within 600 m	29.05	23.04
price	Lowest price per room	180,162.17	107,915.68
time_icn	Time taken to the Incheon Airport	443.61	1332.20
time_gmp	Time taken to Gimpo Airport	1750.53	577.94
time_palac	Time taken to Gyeongbokgung Palace	896.69	427.07
final_score	Sentiment scores	4.09	1.03

presents the list of variables and their descriptive statistics. Hotel-related features, such as the number of rooms, opening year, the lowest nightly rates, were extracted from the country’s Ministry of the Interior and Safety website and Google. These variables were standardized to make them comparable across predictors before applying them to regression models.

3.2. Aspect-Based Sentiment Analysis

This section describes how sentiment scores were estimated from hotel review texts. Figure 3 summarizes the workflow of the sentiment analysis. The process began with preprocessing steps, including the removal of non-ASCII characters, stopwords, and punctuation, along with lowercasing and character normalization. After tokenization, lemmatization was performed using ‘nltk’ and ‘spaCy’ to enhance linguistic accuracy.

The initial dataset included approximately 20,000 online reviews collected from 290 Seoul hotels. Since the data were web-scraped, they included old and new reviews. In order to ensure that the analysis reflects the modern status of hotel service quality and customer experience, only reviews posted from 2021 and onward were retained. This is consistent with the recommendation to prefer temporally relevant information in opinion mining because previous reviews may not represent current service offerings or customer expectations (Liu [23], Thelwall [24]).

We then excluded hotels with fewer than five reviews to ensure the stability and reliability of sentiment averages (Liu [23], Medhat et al. [25]). At the same time, hotels with significantly more reviews could have disproportionately influenced the regression model. Therefore, to maintain fairness and consistency, exactly the same number of reviews per hotel were randomly selected. This approach controls for review count bias and is consistent with best practices in machine learning and econometric modeling, where a balanced number of observations per group helps reduce model variance and overfitting (Harrell [26], Kuhn & Johnson [27]).

Owing to the smaller dataset, a lexicon-based tool for sentiment analysis, TextBlob, was employed rather than deep learning-based models such as BERT. While BERT-based models offer extremely high precision in the case of most natural language processing tasks, they require a lot of computational power and a lot of labeled data to perform optimally (Zhang, Wang, & Liu [28]). On the other hand, TextBlob is lean, easy to use, and has been discovered to work well in identifying sentiment polarity from short texts, as well as product reviews, particularly when interpretability is a concern (Liu, [23], Medhat, Hassan, & Korashy [25]).

To extract relevant features, reviews were parsed into phrases and parts of speech. Key noun phrases such as “location,” “staff,” and “room” were identified and linked to nearby sentiment-bearing adjectives and expressions. The polarity of each aspect phrase was assessed at both the hotel and district levels and then aggregated.

Unlike traditional sentiment analysis that yields a single sentiment score per document, ABSA distinguishes sentiment by aspect, identifying whether particular features are viewed positively, negatively, or neutrally. Reviews were segmented into smaller units and then processed through part-of-speech tagging to isolate aspects (typically noun phrases) from associated sentiment expressions.

While TextBlob handles noun phrase extraction, more advanced systems may use dependency parsing or rule-based methods for higher accuracy. In this study, sentiment for each aspect was computed by analyzing relevant expressions within the sentence or phrase in which the aspect appears. Polarity scores (ranging from −1 for negative to +1 for positive) and subjectivity scores were obtained and aggregated by neighborhood.

These sentiment values were compiled at the hotel level, resulting in a single mean sentiment score per hotel. This score was mapped onto a standardized 1-to-5 scale to align with conventional hotel ratings. The transformed dataset was saved for use in subsequent analysis. Computing sentiment at the hotel level rather than the individual review level mitigates user-level variability, where some guests are consistently more positive or negative than others. Aggregation therefore reduces noise and offers a more stable reflection of overall guest sentiment.

4. Results

4.1. Sentiment Scores

To investigate how hotel and spatial attributes influence guest sentiment, we applied both ordinary least squares (OLS) and spatial autoregressive combined (SAC) models, using sentiment scores derived from text reviews as the dependent variable. The OLS model served as the baseline, while the (SAC) model was employed to account for potential spatial dependencies across hotels.

OLS estimates the linear relationship between a continuous dependent variable and one or more independent variables. The model seeks to minimize the sum of squared residuals between observed and predicted values.

$$Y={\beta }_0+{\beta }_1{X}_1+{\beta }_2{X}_2+\cdots +{\beta }_n{X}_n+\epsilon$$

(1)

where Y is the dependent variable (sentiment score or hotel rating), X_i represents the independent variables such the opening year, number of rooms, distance to airport, bus stop and subway, price, store count, distance to nearest bus stop, convenience store, etc., βi represents the regression coefficients, and ϵ is the error term.

The OLS regression explained approximately 44.85% of the variation in sentiment scores (adjusted R² = 0.448514), indicating a moderately strong model fit (

Table 2. Ordinary least squares for sentiment scores.

Variable	Coefficient	StdError	t-Statistic	p-Value		VIF
Intercept	−0.000000	0.077334	−0.000000	1.000000		--------
open_yr	−0.011127	0.097382	−0.114267	0.909400		1.401114
n_rooms	0.318195	0.094169	3.378982	0.001276	***	1.364701
dist_icn	1.064026	0.468281	2.272195	0.026610	**	52.358468
dist_gmp	−0.932631	0.457273	−2.039553	0.045733	**	50.450570
dist_palac	−0.283181	0.359385	−0.787959	0.433768		20.506012
bus	0.084157	0.124011	0.678625	0.499940		2.264772
subway	0.057737	0.092243	0.625920	0.533702		1.753107
cvs	0.162837	0.112168	1.451722	0.151707	.	2.717527
price	−0.181539	0.150304	−1.207816	0.231780		1.345176
hotel rating	0.614123	0.085413	7.190067	0.000000	***	1.254263
time_icn	−0.895202	0.321161	−2.787394	0.007074	***	148.303715
time_gmp	0.806061	0.310070	2.599606	0.011688	**	145.752485
time_palac	0.016601	0.385984	0.043009	0.965833		23.544065

Note: *** p ≤ 0.01; ** p ≤ 0.05; * p ≤ 0.10; . p ≤ 0.20.

).

The results of ordinary least squares (OLS) regression are important in establishing determinants of guest sentiment for Seoul hotels. Among the predictors employed, number of rooms (n_rooms, β = 0.318, p < 0.01) and hotel rating (hotel_rati, β = 0.614, p < 0.001) were the most notable, validating that highly rated and large hotels experience more favorable guest sentiment. This finding is consistent with previous research, which has shown that hotel size can be a marker for better amenities and service quality (Leung et al. [29]), while star ratings also correlate with customer satisfaction in a strong manner (Ye et al. [30]). Proximity and access-related variables such as distance to Incheon airport, distance to Gimpo airport, time taken to reach Incheon airport, and time taken to Gimpo airport also were found to be statistically significant under uniform standard errors despite registering very high Variance Inflation Factors (VIFs). Usually, a VIF value above 10 suggests multicollinearity (Kutner et al. [31]). These predictors remained significant, however, even with the use of robust standard errors (e.g., time_icn robust p = 0.007), which suggests their explanatory power remains robust. Econometric literature is noted to indicate that although multicollinearity will increase standard errors and reduce the precision of estimates, it will not necessarily produce biased coefficients or render the model useless, particularly if predictors are theoretically and contextually sound (Ken-nedy [32], Wooldridge [33]). More seriously, simply deleting collinear predictors because VIFs are inflated can introduce omitted variable bias and reduce the model’s explanatory power (Dormann et al. [34]). In real spatial applications like tourism or hotel studies, where infrastructural and spatial variables are necessarily correlated, multi-collinearity is typically unavoidable and must be evaluated in relation to theoretical relevance (Anselin [35]).

In addition, to verify the validity of model assumptions, a studentized Breusch–Pagan test for heteroskedasticity was conducted, BP = 16.996, df = 13, p-value = 0.1995. This reveals that the null hypothesis of homoskedasticity cannot be rejected. This verifies the validity of OLS model residual variance and suggests that heteroskedasticity is not a problem in this case (Breusch & Pagan [36]). Hence, the combination of statistically significant predictors, robust error correction, and validation by diagnostic testing ensures the model is methodologically sound. Additionally, when the research intention is predictive and exploratory as opposed to purely causal, as with the norm in spatial hotel sentiment research, retention of collinear but significant variables in the model can ensure the model is more interpretable and applicable (James et al. [37]). Overall, this research underlines the intricate hotel and spatial characteristics that have a strong impact on tourist emotions in urban hospitality environments.

Whereas OLS regression analysis results did reveal several statistically significant predictors of guest sentiment, the model may still be hiding important spatial dynamics. In particular, surrogates like journey time and distance to major transportation nodes had VIFs that were unrealistically high. This reflects the occurrence of strong intercorrelation most probably caused by shared spatial infrastructure, a common characteristic in urbanized zones like Seoul (Dormann et al. [34]). Even when they were statistically significant using robust standard errors, the presence of multicollinearity complicates the interpretation of coefficients and may signal the presence of common spatial processes not properly modeled using the OLS model. Retaining such variables in the model is acceptable if they are both theoretically driven and empirically relevant (Kennedy [32], Wooldridge [33]), but doing so may obscure the presence of spatial spillover effects and induce omitted spatial variable bias (LeSage & Pace [38]).

To explain these problems, the change to a spatial autoregressive combined (SAC) model is justified. Spatial regression techniques enable one to model spatial interdependence explicitly by adding lagged dependent variables, picking up localized influences unobtainable with standard OLS (Anselin, [39] Elhorst [40]). In particular, the SAR model is well-suited to situations where one site’s performance, say, a hotel, is partially determined by near sites, something widely established in tourism and urban studies (Wang & Zhou [12], Valenzuela-Ortiz et al. [41]).

The spatial autoregressive combined model is a form of spatial regression to account for spatial dependence in data, where the value of an observation depends on the values of nearby observations. The model encompasses both the typical predictors influencing the dependent variable as well as other values of the dependent variables at near points weighted by a spatial weight matrix.

Anselin [39] suggested the spatial lag model as a solution to the common observation independence violation in geographic data. The spatial lag model directly addresses spatial autocorrelation in the regression equation and produces less biased and more accurate estimates in spatially structured data.

$$\mathrm{Y}=\rho \mathrm{W}\mathrm{Y}+\mathrm{X}\mathsf{\beta }+{\left(\mathrm{I}-\mathsf{\lambda }\mathrm{W}\right)}^{-1}\epsilon ,$$

(2)

where $\mathrm{Y}$ denotes the vector of the dependent variable, such as hotel sentiment scores, while X represents the matrix of explanatory variables with β as the corresponding vector of regression coefficients. The term $\rho \mathrm{W}\mathrm{Y}$ captures the spatial lag component, where $\rho$ is the spatial autoregressive coefficient and $\mathrm{W}\mathrm{Y}$ is the weighted average of neighboring values of the dependent variable, based on the spatial weights matrix W that defines the neighborhood structure (e.g., based on distance or contiguity). Beyond the spatial lag, the SAC model also introduces spatial dependence in the residuals through the term ${\left(\mathrm{I}-\mathsf{\lambda }\mathrm{W}\right)}^{-1}\epsilon$, where λ is the spatial error coefficient and ε represents the independently and identically distributed error terms.

To construct the spatial weights matrix, we employed the k-nearest neighbors (KNN) method with k = 4 as the adjacency criterion. Given the highly irregular spatial distribution of the hotels, many of which are clustered in specific areas, a fixed-distance approach would have resulted in numerous spatial isolates (islands). This would have undermined the connectivity required for spatial regression analysis, rendering it impractical. Therefore, the KNN method was chosen to ensure that each hotel maintained a minimum level of spatial interaction with its neighbors.

The results of the spatial autoregressive combined analysis confirm the presence of spatial dependence in hotel sentiment scores across Seoul. Both the Robust LM Error (p = 0.021) and Robust LM Lag (p = 0.037) tests are statistically significant (

Table 3. Spatial autoregressive combined model for sentiment scores.

Variable	Coefficient	StdError	z-Statistic	p-Value
Intercept	−0.011014	0.037730	−0.291917	0.770350
open_yr	−0.034031	0.090731	−0.375072	0.707607
n_rooms	0.296977	0.077913	3.811663	0.000138	***
dist_icn	0.926400	0.388196	2.386421	0.017013	**
dist_gmp	−0.824747	0.364829	−2.260642	0.023781	**
dist_palac	−0.073695	0.216523	−0.340358	0.733587
bus_600	0.016460	0.094790	0.173648	0.862142
subwy_600	0.038380	0.080415	0.477271	0.633169
cvs_600	0.135497	0.073774	1.836663	0.066260	*
price	−0.156782	0.142281	−1.101923	0.270495
hotel rating	0.574012	0.083361	6.885838	0.000000	***
time_icn	−0.773451	0.316124	−2.446675	0.014418	**
time_gmp	0.685585	0.310249	2.209790	0.027120	**
time_palac	−0.123854	0.242348	−0.511061	0.609309
lag y (rho)	0.444165	0.141232	3.144936	0.001661	***
Lag residual (lambda)	−0.852893	0.340537	−2.504553	0.012261	**

Note: *** p ≤ 0.01; ** p ≤ 0.05; * p ≤ 0.10; . p ≤ 0.20.

) suggesting that sentiment at a given hotel is positively influenced by the sentiment levels of nearby hotels. This supports prior findings that tourist experiences are spatially correlated due to shared neighborhood conditions and environmental cues (Wang & Zhou [12], Valenzuela-Ortiz et al. [41]).

Of the explanatory variables used, number of rooms (β = 0.297, p < 0.001) and overall rating of a hotel (β = 0.574, p < 0.001) were the most predictive in our research. This empirical finding is consistent with existing hospitality literature indicating that service capacity (e.g., the size of the hotel or room number) and perceived quality of service are important drivers of guest satisfaction (Kim & Kim [42], Susanto et al. [43]). Additionally, spatial accessibility indicators such as travel distance to Incheon Airport (β = 0.926, p = 0.017) and travel time to Incheon (β = –0.773, p = 0.014) are significant, again highlighting the twin relevance of proximity as well as travel time in shaping hotel perceptions, an assertion repeated in work on transport-based accessibility and service industry performance in Seoul (Song et al. [44], Latinopoulos [2]).

The inclusion of convenience stores within 600 m (β = 0.135, p = 0.066) shows significance too, suggesting that neighborhood commercial density also contributes to improved guest sentiment, although with a smaller effect size. Interestingly, traditional location variables such as distance to palaces or presence of subway stations did not show significant effects, reflecting Seoul’s uniformly high accessibility baseline and the nuanced preferences of urban tourists (Kim & Cheoi [45]). Overall, the significant spatial lag term and localized effects underscore the necessity of spatial modeling for understanding variation in hotel sentiment.

4.2. Hotel Ratings

Building on the results from the sentiment model, we also applied ordinary least squares (OLS) regression to analyze hotel ratings as the dependent variable. While sentiment scores reflect emotional responses extracted from textual reviews, formal ratings offer a more structured and standardized measure of guest satisfaction. This comparison allows us to examine whether similar factors influence both types of evaluations.

As shown in

Table 4. Ordinary least squares for hotel ratings.

Variable	Coefficient	StdError	t-Statistic	p-Value		VIF
intercept	0.000000	0.102414	0.000000	1.000000		--------
open_yr	0.300536	0.112851	2.663122	0.009850	**	1.290889
n_rooms	0.095244	0.118956	0.800665	0.426379		1.353323
dist_icn	0.388319	0.743406	0.522351	0.603290		52.169336
dist_gmp	−0.712732	0.698658	−1.020144	0.311621		49.813422
dist_palac	0.739753	0.477086	1.550564	0.126100	.	19.819636
bus_600	−0.262651	0.140732	−1.866321	0.066730	.	2.178246
subwy_600	0.067333	0.150384	0.447744	0.655902		1.747420
cvs_600	−0.214063	0.189763	−1.128058	0.263640		2.660829
price_0617	0.154292	0.106181	1.453111	0.151240		1.315317
time_icn	0.067917	0.686823	0.098886	0.921546		148.297929
time_gmp	0.124235	0.661658	0.187763	0.851676		145.733126
time_palac	−0.780164	0.528366	−1.476561	0.144859	.	22.780650

Note: *** p ≤ 0.01; ** p ≤ 0.05; * p ≤ 0.10; . p ≤ 0.20.

, the hotel rating model yielded a lower adjusted R-squared value of 0.0484, indicating that it explained roughly 0.4% of the variation in ratings. This is considerably less than the sentiment model and suggests that formal ratings may be influenced by a narrower range of factors or exhibit less variability.

The OLS regression model for hotel rating determinants presents fewer statistically significant predictors than the sentiment score model, suggesting varied forces underlying how travelers rate their hotel stay as opposed to how they feel in responding comments. Hotel opening year is determined to have statistically significant and positive impact on hotel ratings at the 1% level in the robust model (β = 0.301, p = 0.0099) and indicates that newer hotels are given higher ratings. This is also aligned with past research proving that recency of construction or renovation is normally followed by better maintenance, newer facilities, and newer services, all of which positively affect perceived quality (Xie & Zhang [46]). Otherwise, the number of rooms is not significant here (p = 0.43), whereas it was a strong predictor of sentiment in the initial model (p = 0.0013), and it is predictive of the observation that guests might see large hotel size as having more influence on experience compared to rating.

Although distance and travel time to Gyeongbokgung Palace were only marginally significant (p = 0.13 for distance; p = 0.145 for travel time), the relatively large coefficients (β = 0.74 and β = –0.78) suggest that spatial accessibility may exert a subtle influence on how guests assign formal ratings to hotels. Interestingly, bus stop accessibility within 600 m shows a borderline negative association with hotel ratings (β = –0.26, p = 0.067). One possible explanation is that extremely high proximity to bus stops may introduce noise or reduce perceived exclusivity, especially in premium hotel settings. This idea aligns with urban planning literature showing that while transit access is often beneficial, proximity can sometimes carry a “penalty” due to overcrowding or nuisance effects, particularly in dense urban areas (Hao & Peng [47]).

Importantly, the lack of significance for variables such as distance to Gyeongbokgung Palace and subway stations supports the argument that Seoul’s compact urban structure and extensive public transportation network may reduce the explanatory power of spatial proximity when it comes to guest evaluations (Latinopoulos [2]). This supports prior work suggesting that hotel ratings, unlike emotional sentiment, are more influenced by cognitive judgments tied to service quality, cleanliness, or amenities, rather than logistical convenience (Yang et al. [48]).

Model diagnostics (Koenker p = 0.55; Jarque–Bera p = 0.78) indicate that the OLS assumptions are satisfied, yet the insignificance of most predictors highlights that formal ratings may be shaped more by latent factors, such as brand reputation, service experience, or reviewer-specific biases, that are not captured in this model (Vargas-Calderón et al. [49]).

5. Discussion

This study analyzed the determinants of hotel satisfaction in Seoul through ordinary least squares (OLS) and spatial autoregressive combined (SAC) models, focusing on two dependent variables: sentiment scores derived from guest reviews and formal hotel star ratings. The results revealed important distinctions in how spatial and hotel attributes influence these two satisfaction metrics.

Under the OLS model to predict sentiment scores, several variables were statistically significant. Of particular note, hotel rating and number of rooms both had strong positive effects on sentiment, suggesting greater hotels and high-star hotels tend to have more positive guest perception. These findings are in line with previous research that accounts for hotel size and official rating to guest satisfaction in terms of greater resources, service quality, and perceived professionalism (Leung et al. [29], Xie & Zhang [46]). Spatial accessibility measures, both travel time and distance, to major airports (Incheon and Gimpo) also showed statistical significance, identifying the twin role of geographic proximity and perceived convenience in shaping guest sentiment. These results support previous evidence that accessibility is a crucial factor in tourism competitiveness and hotel city performance (Song et al. [44]).

The SAC model provided more evidence by confirming the presence of spatial dependence in hotel sentiment. The spatial lag coefficient $\left(\rho \right)$ was significant and positive and indicated that sentiment in one hotel is partially explained by the sentiment in nearby hotels. Such geographical spillover effect corroborates earlier findings of research that hotel experience is not only driven by individual hotel amenities but also by shared neighborhood characteristics that encompass security, hygiene, or vibrancy (Wang & Zhou [12], Valenzuela-Ortiz et al. [41]). Travel duration to Incheon airport continued to matter and was adverse in SAC results, pointing out that accessibility by time, as opposed to raw distance, is an essential driver of visitor ratings, a trend consistent with Seoul’s emphasis on multi-modal transportation infrastructure (Lee et al. [50]).

Conversely, the OLS hotel rating model produced fewer significant variables and a lower adjusted R² that proves star ratings are the outcome of a different evaluation criterion set. For this instance, opening year (β = 0.301, p < 0.01 robust) was significant, and what this suggests is that more recent hotels possess slightly higher official scores. This may be due to the ability of new hotels to meet newer standards of design, service technologies, and customer expectations (Xie & Zhang [46]). Surprisingly, variables like the number of rooms and spatial accessibility, important in the sentiment model, were not good predictors of formal ratings. This aligns with the position that written emotions better pick up more experiential and context-dependent details, whereas numerical ratings better pick up generalized or procedural aspects of hotel evaluation (Li et al. [51], Zhang et al. [52]).

The contrast between hotel rating and sentiment models reveals a more general observation: neighborhood and spatial effects are more significant in modeling qualitative guest experiences than in estimating standardized hotel ratings. This suggests the importance of including spatial regression techniques when examining perceptual outcomes in tourism studies. This research further shows that spatially auto-correlated models such as SAC can reveal relationships masked by multicollinearity or confounded urban characteristics in conventional OLS specifications.

Together, these findings have several theoretical implications. First, the comparison between hotel rating and sentiment models accentuates the multidimensionality of tourist satisfaction. Expectation–Disconfirmation Theory (Oliver [3]) assists us in this regard: formal ratings may align with baseline service expectations, whereas sentiment scores trace emotional disconfirmation as a reaction to spatial context and neighborhood experience. Second, the prevalence of widespread spatial spillovers affirms the argument that guest satisfaction is not an isolated outcome but one latent within the spatial fabric of the city, extrapolating Tobler’s Law to the hospitality domain. Such discoveries prompt us to hypothesize tourist satisfaction not just as a function of service delivery but as an effect of urban spatial form and neighborhood arrangements.

This study also offers serious methodological contributions. The difference between OLS and SAC models underscores the limitations of not taking into account spatial dependence in hotel studies. Standard regression models suppress inter-dependencies between hotels, whereas spatial models accentuate neighborhood spillovers and provide richer insights into patterns of sentiment. The use of buffer-based Euclidean distances in the current study also demonstrates the value of a practical, transparent, and commonly applied accessibility measure. The 600 m buffer matches common definitions of walkability from urban and tourism literature and provides us with a reliable baseline to make cross-place comparisons. To do so, though, this approach suggests an important line of future research: enhancing accessibility measures to more accurately capture the more subtle character of urban mobility and tourist experience.

Generally, this study confirms that spatial and neighborhood impacts are more significant influences on qualitative guest experiences than in explaining standardized ratings. Theoretically, it contributes to hospitality research by situating tourist satisfaction within broader spatial and urban contexts. Methodologically, it highlights the virtue of the integration between sentiment analysis and spatial econometrics and suggests next-generation approaches to incorporating route-based, perceptual, and network-based accessibility models. By placing hotel satisfaction in the geographical framework, this research explains the dynamics of urban hospitality.

6. Conclusions

This study finds that sentiment scores are more responsive to spatial characteristics, such as proximity to transit, airports, and landmarks, than formal hotel ratings. Emotional responses reflect not only service quality, but also the broader urban environment in which a hotel is situated. By contrast, star ratings appear more closely tied to internal hotel attributes like cleanliness and service.

For example, guests may leave positive ratings for service but express dissatisfaction with noise or congestion in their reviews. This divergence highlights the value of analyzing sentiment alongside ratings to better understand how location contributes to guest satisfaction.

These insights have practical implications. Hotels in high-density or noisy areas may benefit from investing in soundproofing or wellness amenities, while quieter or residential hotels could market relaxation and calm. For urban planners, understanding how spatial form and transit infrastructure affect tourist sentiment can inform decisions about tourism development and public investment.

By combining sentiment analysis with spatial data, this study contributes to a more comprehensive understanding of urban hospitality experiences. The findings may support hotel managers in refining services and targeting marketing strategies, and guide policymakers in enhancing the tourist experience through better infrastructure and land use planning.

However, several limitations remain. The use of Euclidean distance does not fully capture real-world travel behavior, which may affect accessibility measures. The study is also limited to reviews from TripAdvisor, which may not represent domestic tourists who use local platforms. Furthermore, the analysis is cross-sectional and does not account for seasonal or event-driven fluctuations.

Future research should incorporate network-based travel distances, time-sensitive data, additional review platforms, and deep learning models (e.g., BERT) to improve sentiment detection and generalizability across different urban contexts.