📈 economics

Spatial Disparities in Short-term Rental Guest-Perceived Accommodation Quality: A Twelve-City Comparative Analysis Using Airbnb Review Data and Explainable Machine Learning

This study employs an explainable machine learning pipeline on 297,227 Airbnb reviews across twelve global cities to quantify spatial disparities in guest-perceived accommodation quality using a novel inequality index, revealing that moderate-restrictive regulatory regimes exhibit higher quality dispersion than permissive or highly restrictive ones while identifying Superhost status as the dominant, location-conditioned predictor of satisfaction.

Original authors: Jing Gao, Dong Lin, Jianfeng Hu

Published 2026-08-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jing Gao, Dong Lin, Jianfeng Hu

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to understand how happy people are with their homes in a giant, bustling city. Usually, to get this answer, researchers have to knock on thousands of doors, ask long surveys, and wait years for the results to come back. It's like trying to map a storm by asking people what the weather was like last month. But in the digital age, there's a new way to listen: the "digital footprint." Every time someone stays at a vacation rental, they often leave a review. These reviews are like a massive, real-time diary of how people feel about their temporary homes, written by millions of guests. This paper lives in the world of urban geography and machine learning, two fields that study where things happen and how computers can find hidden patterns in huge piles of data. The big question here is simple but tricky: Does the quality of these vacation rentals vary wildly from neighborhood to neighborhood, and does the city's rules about renting out homes change that? Understanding this matters because if some areas are full of great stays while others are full of bad ones, it tells us something about how fair and well-run a city's housing services are.

Now, let's dive into what the researchers actually did. They treated Airbnb reviews not just as star ratings, but as a direct measure of "guest-perceived accommodation quality." Think of it like a giant, global taste test where millions of people are rating the "flavor" of their stay. The team gathered data from 297,227 listings across 12 cities on four different continents, including places like Amsterdam, Barcelona, London, and Bangkok. They wanted to see if the strictness of a city's rules (like banning rentals or limiting how many days you can rent) changed how uneven the quality of these rentals was.

To crack this code, they built a super-smart computer model called XGBoost. Imagine this model as a detective that looks at clues like the price, the type of room, and whether the host is a "Superhost" (a top-rated, experienced host) to guess if a guest will be happy. The model was incredibly good at its job, correctly predicting high satisfaction 86.7% of the time. But the real magic wasn't just in the guessing; it was in the "explainability." The researchers used a tool called SHAP to ask the model, "Why did you make that guess?" The answer was clear: the Superhost status was the biggest clue, accounting for 62.2% of the model's decision-making power. However, being a Superhost isn't just about the person; it's also about where they live. These top hosts tend to cluster in the city centers, meaning the "who" is often tied to the "where."

The researchers then created a special score called the STR-GPAQ Inequality Index (SII). Think of this as a "fairness meter" for vacation rentals. A low score means everyone gets a pretty similar, good experience. A high score means the experience is a rollercoaster—some people get amazing stays, while others get terrible ones.

Here is where the plot thickens. The study found a surprising pattern in how city rules relate to this fairness meter.

Amsterdam, which has very strict rules (like a 30-day cap and needing a permit), had the lowest inequality (SII = 0.031) and the highest satisfaction rate (94.6%). It was like a well-organized library where every book is in good condition.
Barcelona, which has "moderate-restrictive" rules (a mix of limits and freezes on licenses), had the highest inequality (SII = 0.115) and the lowest satisfaction rate (76.8%). This was more like a chaotic bazaar where you might find a treasure or a broken toy.
Permissive cities (like Los Angeles and Bangkok) and highly restrictive cities (like Paris) generally fell somewhere in the middle, with moderate inequality.

The authors suggest that this "moderate-restrictive" middle ground might accidentally create a split market, leading to more uneven quality, but they are careful to say this is a descriptive association, not a proven cause-and-effect. They explicitly state that their study is a snapshot in time and cannot prove that the rules caused the inequality; other factors like tourism intensity or housing markets could be to blame.

One crucial detail the paper rules out is that this tool is perfect for finding the worst rentals. The model was great at spotting the good ones but struggled to identify the bad ones (it only caught 4.2% of the low-quality listings). The authors admit this means their "fairness meter" is best used to measure how spread out the good experiences are, rather than as an alarm system to hunt down bad apartments.

In the end, this paper offers a new, playful, and powerful way to look at city housing. Instead of waiting for slow surveys, we can listen to the digital chatter of millions of guests. The findings suggest that while top hosts drive quality, the city's rules might shape how evenly that quality is spread. The authors propose this method as a "proof-of-concept"—a working prototype that city planners could use to periodically check if their housing services are fair, while acknowledging that more research is needed to understand the full story behind the numbers.

Technical Summary: Spatial Disparities in Short-term Rental Guest-Perceived Accommodation Quality

Problem Statement
Traditional methods for assessing housing satisfaction, primarily household surveys, suffer from temporal lags, high resource costs, and a lack of fine-grained spatial detail necessary for neighborhood-level policy diagnostics. This data gap is particularly acute in short-term rental (STR) markets, where rapid platform growth (e.g., Airbnb) has altered urban housing dynamics, yet regulatory policy often relies on broad aggregate data that fails to capture detailed spatial variations in service quality. The study addresses the need for a direct, systematically measurable, and spatially explicit indicator of STR service quality to diagnose housing service deprivation across diverse urban contexts.

Methodology
The study employs an integrated explainable machine learning pipeline applied to a dataset of 297,227 Airbnb listings across twelve cities on four continents (North America, Europe, Asia, and Oceania). The cities were selected via stratified purposive sampling to represent diverse regulatory regimes (permissive, moderate-restrictive, and highly restrictive) and maximum continental representation within the available data, explicitly noting the absence of African and South American cities.

Data and Variables:
- Source: Publicly available anonymized data from Inside Airbnb (snapshot from June 2026).
- Dependent Variable: A binary classification of Guest-Perceived Accommodation Quality (STR-GPAQ). Listings with review scores $\ge$ 4.5 are classified as "High Satisfaction," while those < 4.5 are "Non-High."
- Independent Variables: Geographic (obfuscated coordinates, distance to center), Property (price, room type, capacity), Host (Superhost status, review counts), and Institutional (city dummy variables). Variables with high missingness (e.g., cleaning fees) were excluded.
Model Architecture:
- Algorithm: XGBoost binary classification.
- Validation: Spatial block cross-validation (9 blocks based on latitude/longitude quantiles) was used to prevent data leakage from spatial autocorrelation, alongside standard random k-fold validation.
- Interpretability: SHAP (SHapley Additive exPlanations) values were used to decompose model predictions and quantify feature importance and interactions.
The STR-GPAQ Inequality Index (SII):
- A novel metric adapted from the Gini coefficient, applied to the predicted probabilities of high satisfaction rather than raw scores.
- Formula: $SII_c = (1 / (2 \bar{P}_c N_c^2)) \sum_i \sum_j |P_i - P_j|$ , where $P_i$ is the predicted probability for listing $i$ .
- Purpose: To measure the distributional inequality of high-quality accommodation across listings within a city. It is a scalar measure of dispersion, not a direct measure of spatial clustering, though its spatial validity is supported by Moran's I diagnostics and grid-sensitivity analysis.

Key Results

Model Performance: The XGBoost model achieved a cross-validated accuracy of 0.867, an AUC of 0.789, and an F1-score of 0.928. However, due to extreme class imbalance (positive class >85%), the negative class recall was very low (0.042). The authors note the model is optimized to identify structural drivers of inequality rather than to detect individual low-quality listings (early warning).
Feature Importance: The "Host Superhost" status emerged as the dominant predictor, accounting for 62.2% of feature importance. However, SHAP interaction analysis revealed that this effect is spatially conditioned, being significantly stronger in central districts with high Point of Interest (POI) density compared to peripheral areas. City dummy variables collectively accounted for ~25% of importance, indicating that baseline norms and platform maturity vary significantly by location.
Regulatory Associations: A descriptive pattern emerged linking regulatory regimes to inequality:
- Highly Restrictive (e.g., Amsterdam): Lowest SII (0.031) and highest satisfaction rate (94.6%).
- Permissive (e.g., Los Angeles, Bangkok): Moderate SII values (0.057–0.083).
- Moderate-Restrictive (e.g., London, Barcelona): Highest SII values. Barcelona exhibited the highest inequality (SII = 0.115) and lowest satisfaction (76.8%).
Robustness: SII rankings remained stable across spatial aggregation scales (0.5 km to 2.0 km grids) and were confirmed to be distinct from random spatial baselines. Ordered logistic regression confirmed results were robust to the 4.5 threshold choice.

Key Contributions

Conceptual Redefinition: The study redefines Airbnb guest reviews as a direct, measurable construct of STR-GPAQ, positioning it as a complementary (though not synonymous) source for measuring housing service deprivation, with explicit boundaries on construct validity.
Methodological Framework: It introduces and validates the STR-GPAQ Inequality Index (SII), a Gini-inspired metric designed specifically for the STR domain, demonstrating its stability across spatial scales and its ability to facilitate cross-city comparisons.
Cross-City Comparative Design: By analyzing twelve cities across four continents with varying regulatory regimes, the study provides a descriptive benchmark showing that moderate-restrictive regimes may be associated with higher distributional inequality than either permissive or highly restrictive regimes.
Replicable Pipeline: The authors provide a complete analytical framework (XGBoost + SHAP + SII + Spatial Diagnostics) that can be adapted by researchers and practitioners for periodic monitoring of service deprivation in other domains.

Significance and Claims
The paper positions itself primarily as a methodological proof-of-concept rather than a causal study. Its significance lies in:

Diagnostic Utility: Offering a scalable, data-driven approach to visualize and quantify spatial disparities in STR service quality using digital trace data, overcoming the limitations of traditional surveys.
Policy Benchmarking: Providing a descriptive baseline (the SII rankings) that future quasi-experimental studies can use to assess the causal effects of specific regulatory interventions.
Transparency: Emphasizing explainability (via SHAP) and construct validity boundaries to mitigate the risks of algorithmic bias and the misinterpretation of platform data as a proxy for resident welfare.

The authors explicitly caution that their cross-sectional design precludes causal attribution. The observed association between moderate-restrictive regimes and higher inequality is presented as a hypothesis-generating finding that requires panel data or quasi-experimental validation (e.g., difference-in-differences) to confirm causality and rule out confounding factors like tourism intensity or housing market tightness. Furthermore, the study acknowledges that STR-GPAQ reflects guest satisfaction (transient, service-oriented) and is not synonymous with long-term resident housing satisfaction.

Technical Summary: Spatial Disparities in Short-term Rental Guest-Perceived Accommodation Quality

More like this