Integrating Machine Learning, SHAP Interpretability, and Deep Learning Approaches in the Study of Environmental and Economic Factors: A Case Study of Residential Segregation in Las Vegas

Abstract

Over the past two decades, research on residential segregation and environmental justice has evolved from spatial assimilation models to include class theory and social stratification. This study leverages recent advances in machine learning to examine how environmental, economic, and demographic factors contribute to ethnic segregation, using Las Vegas as a case study with broader urban relevance. By integrating traditional econometric techniques with machine learning and deep learning models, the study investigates (1) the correlation between housing prices, environmental quality, and segregation; (2) the differentiated impacts on various ethnic groups; and (3) the comparative effectiveness of predictive models. Among the tested algorithms, LGBM (Light Gradient Boosting) delivered the highest predictive accuracy and robustness. To improve model transparency, the SHAP (SHapley Additive exPlanations) method was employed, identifying key variables influencing segregation outcomes. This interpretability framework helps clarify variable importance and interaction effects. The findings reveal that housing prices and poor environmental quality disproportionately affect minority populations, with distinct patterns across different ethnic groups, which may reinforce these groups’ spatial and economic marginalization. These effects contribute to persistent urban inequalities that manifest themselves in racial segregation and unequal environmental burdens. The methodology of this study is generalizable, offering a reproducible framework for future segregation studies in other cities and informing equitable urban planning and environmental policy.

1. Introduction

Over the years, residential segregation has been an unavoidable topic in addressing racial discrimination issues in American cities [1,2]. During this time, as the focus of urban development has shifted, various theories explaining the reasons for residential segregation have emerged [3]. Early theoretical models primarily discussed the spatial forms of segregation, such as the Concentric Zones Model based on urban economic development, the Sectors Theory triggered by traffic conditions [1,4], and the Multiple Nuclear Model advocated by Harris and Ullman [5]. Subsequently, quantified models describing residential segregation gradually matured, each with their own theoretical inclination [6]. The spatial assimilation model explains how differences in socioeconomic status and cultural adaptability among races lead to segregation [7]. The place stratification model focuses on the persistence of intergroup relationship biases [8,9]. The widely used Schelling’s Model is based on the residential preference theory [10,11]. Subsequently, derived theories such as Class theory [12] and Self-Segregation theory [13] have also emerged. While numerous theories are attempting to explain racial housing distribution disparities, some critical issues remain unsolved. By utilizing Las Vegas as a research site, this study aims to address the following shortcomings in the field of residential segregation

1.1. Insufficient Research on the Impact of Housing Prices on Residential Segregation

After the enactment of the Civil Rights Act of 1964 and the Fair Housing Act of 1968, some African Americans and other minorities, who had improved their educational levels and economic conditions, still faced barriers to overcoming residential segregation [14,15]. Even after socioeconomic advancement, many minority households continued to encounter obstacles in accessing desirable neighborhoods due to entrenched racial discrimination in the housing market. Discriminatory practices, such as racial steering and the risk of property devaluation triggered by changes in racial composition, interact with systemic poverty to form a feedback loop. This loop perpetuates residential segregation by restricting the ability of minority households to accumulate wealth and relocate across neighborhoods [15,16]. Recent evidence also indicates that pioneering black families who were among the first to move into white neighborhoods faced the risk of property devaluation due to changes in racial composition, despite bearing the burden of high housing prices [16]. Therefore, changing racial biases in the real estate market is crucial for improving the socioeconomic status of minority communities and thereby alleviating residential segregation [17,18].

Las Vegas experienced a period of rapid expansion characterized by a more typical unregulated housing market than most American cities [19,20]. Meanwhile, discriminatory practices such as redlining, restrictive covenants, and racial steering were prevalent in the real estate industry of Las Vegas. These practices systematically excluded people of color, particularly African Americans and Latinos, from certain neighborhoods and limited their access to housing opportunities [21,22]. The ghettos of Las Vegas starkly contrast with the thriving entertainment venues in the area, and this disparity remains difficult to address even to this day. Therefore, for cities like Las Vegas, where the development of the modern real estate industry is closely related to racial residential segregation, it is still necessary to consider housing prices as the core factor in building residential segregation models [23,24].

While earlier frameworks such as Schelling’s Model do not incorporate external factors such as housing prices, socioeconomic conditions, or historical environmental dynamics—making them less suited to address specific societal inequalities—subsequent theories have emphasized more complex relationships [24]. Notably, theories of housing sub-markets and filtering conceptualize the housing market as a series of overlapping sub-markets stratified by location, price, and tenure. These sub-markets reflect and reinforce socioeconomic and racial hierarchies, shaping residential accessibility across groups [25,26]. The filtering process—where higher-income households vacate older or lower-quality units that then become available to lower-income residents—can entrench patterns of segregation over time. In addition, local housing markets often operate in disequilibrium, where supply–demand mismatches, regional planning, and historical zoning all affect spatial outcomes [27].

Closely related to this is the concept of residential income segregation, wherein income inequality fuels spatial division as households sort themselves by economic status. Reardon and Bischoff demonstrate how income-correlated residential preferences result in spatial clustering by income level, often aligning with racial demographics in U.S. cities. Such preferences, when coupled with structural market constraints and legacies of discrimination, reinforce both racial and income-based segregation [28].

While empirical studies such as Mikulas et al., Brasington et al., and Box-Couillard and Christensen have incorporated housing prices into segregation analyses at various scales, these models often remain limited to aggregate-level associations [29,30,31]. Our study builds on this body of work by constructing a street-scale predictive model that integrates housing price dynamics with micro-scale environmental and perceptual data using both econometric and deep learning methods. This integrated approach captures the complex interplay between environmental inequities, economic sorting, and racialized spatial structures [29,30,31].

1.2. Influence of Local Environmental Factors on Residential Segregation

In Las Vegas, environmental justice issues are not merely outcomes of residential segregation—they also act as catalysts that reinforce spatial separation. Unequal access to local environmental amenities, such as green space, tree coverage, and clean air, significantly influences neighborhood desirability and residential decision-making. Consequently, environmental inequities—such as exposure to pollutants, lack of tree canopy, or elevated heat levels—disproportionately burden minority and low-income populations, exacerbating both health outcomes and property devaluation. In Las Vegas, the desert climate amplifies these disparities, where vulnerable communities face higher cooling costs and inadequate green infrastructure.

Situated in a desert region, vulnerable communities reside in urban areas with harsher temperatures and higher cooling costs [32]. The unique climate has led residents, particularly those in high-crime neighborhoods, to use large, sturdy fences to deter crime and protect against dust storms, while these structures provide short-term protection, they can also reinforce physical and psychological boundaries between communities, exacerbating social separation [33]. Furthermore, recent discussions on environmental justice have been sparked by air pollution issues from some factories and military facilities in the Las Vegas area [34]. As a result, environmental injustice contributes to a cycle of disinvestment, health disparities, and declining property values—making these areas less desirable and further segregating them from more affluent, environmentally secure neighborhoods. In summary, the housing prices, geographical climate characteristics, community physical environmental structures [35,36], and local immigration history collectively contribute to the distinct residential segregation issues in Las Vegas. Naturally, minority residents in this area are more likely to face environmental justice issues such as air pollution, heat island effects, and lack of green spaces [37,38], factors that deepen the degree of racial residential segregation to varying degrees.

To better understand the mechanism of residential segregation in Las Vegas, this study will establish a model that comprehensively utilizes economic environmental elements to explain and predict racial distribution. A variety of multilevel economic environmental factors will be integrated into the model, including housing prices, street view, POI data, as well as environmental justice indicators such as land surface temperature, air quality, and vegetation density. This racially segregated prediction model, grounded in the reality of Las Vegas, may provide local policymakers with more comprehensive planning and design guidelines to address housing inequality issues.

1.3. Research Scale and Multiracial Subject Issues in Residential Segregation Studies

Some analytical models in previous studies have failed to capture the residential segregation situations of multiple races. For instance, Schelling’s Model simplifies race into a binary category, unable to depict the complex characteristics of various racial groups [39,40]. As multiple minority ethnic groups, including Hispanic and Asian populations, face their own residential segregation challenges, this binary model no longer meets the needs of research on racially diverse American cities [41,42,43]. Similar situations have also occurred in Las Vegas [44]. Therefore, this paper will meticulously discuss the residential segregation issues faced by various minority ethnic groups to uncover the similarities and differences among races in this regard.

Currently, there is a dearth of street-scale studies, which are most relevant to the residential environment, in the field of residential segregation research. The choice of research scale is crucial for the analysis of racial residential segregation outcomes [45]. The majority of studies use census tracts for analysis, or similar community scales [46], and block group-level analyses [47], with a few using street-level analyses [48] or conducting multi-scale comparative studies [49]. Therefore, this study will utilize POI data and street view data to describe the urban economic environment characteristics of Las Vegas. We attempt to unify the data of each variable at the street scale to more accurately reflect the spatial scale at which residential segregation occurs.

1.4. Improve the Accuracy of Residential Segregation Models

Previous research has been relatively lacking in an exploration into the causal relationships among variables in residential segregation models. Early studies often relied on common sense to establish postulated causal models [50]. However, this approach has drawbacks as it depends on judgments based on prior knowledge and may not be effective in exploring models that cannot capture causal relationships [51]. Similar issues of causal ambiguity are widespread in observational data studies across various social sciences [52,53]. Therefore, this paper will employ the Treatment Agnostic Regression Network (TARNet) method proposed by Shali et al. to infer causality among variables. TARNet is a neural network-based method used for estimating individual treatment effects in causal inference. This method effectively addresses interference and endogeneity issues in observational data, preventing model overfitting, and thereby enhances the performance of subsequent predictive models [54].

Furthermore, optimization algorithms such as machine learning methods can be employed to enhance the accuracy of predictive models. Some scholars have already attempted to utilize machine learning methods and social justice data for predicting residential segregation. For instance, the feasibility of using COVID-19 mortality data for predicting residential segregation has been validated [55]. Progress has also been made in research related to using machine learning to predict racial segregation maps [56]. However, no study has yet explored whether machine learning methods can be used to leverage housing prices for predicting residential segregation. To improve the accuracy of residential segregation models, this study will comprehensively integrate traditional econometric models, machine learning, and deep learning methods. Various algorithms will be employed to compare and investigate the optimal machine learning algorithm for predicting residential segregation using housing prices.

In previous studies, the SHAP method was adopted to interpret the regression model, in order to explain the relationship between different factors and urban vitality [57,58,59], as SHAP assigns a prediction-specific value to each feature, which helps in interpreting the predictions. Therefore, it ensured feature consistency and model stability [60].

By filling in the gaps of the previous studies while focusing on the three following questions, this article attempts to construct a predictive model to explain the relationships between housing price, environmental factors, and racial residential segregation.

RQ1: Is there a correlation between economic environmental factors, primarily housing prices, and residential segregation? RQ2: How do housing prices and environmental factors influence residential segregation for different ethnic groups? RQ3: Which machine learning method is most suitable for predicting residential segregation in the context of this study? How can we interpret the relationship between housing prices, environmental factors, and residential segregation in the Las Vegas area?

The remainder of this paper is organized as follows. Section 2 describes the various quantitative methods utilized in this paper and the specific problems addressed by each method. Section 3 will present the data analysis results of each process. In this chapter, we first obtain descriptive statistics for each variable, particularly examining the normality and correlation of the variables. Additionally, the Multiple Ordinary Least Squares method is employed to examine the relationship between socioeconomic environmental factors and population distribution among different racial groups. Before the prediction model is employed, we use the TARNet method to explore the causal relationships between variables, which ensures we can obtain valid results from further prediction. In the final part of Section 3, to achieve the best model performance, we compared various machine learning algorithms and utilized the SHAP interpreter to elucidate the model mechanisms between economic environmental factors and residential segregation. The final research conclusions and limitations will be stated in Section 4, Section 5 and Section 6.

2. Methodology

This study adopts a comprehensive methodological framework that integrates traditional econometric analysis and machine learning models to be applied to a multi-dimensional dataset to explore how environmental factors and housing prices jointly affect residential segregation (Figure 1).

The study commenced by collecting housing price data for the Las Vegas area from Zillow using a custom-built web scraper. Additionally, points of interest (POI) data were retrieved via the OpenStreetMap API to serve as economic and commercial indicators. Concurrently, environmental and demographic data were sourced from the Centers for Disease Control and Prevention (CDC), including specifics on green spaces and air quality from environmental data interfaces. Street-level environmental perception data were also extracted from images sourced from Google Street View. This diverse data collection strategy constitutes a multidimensional dataset composed of macro-level demographic indicators and micro-scale environmental and economic features.

The study uses several variables from the CDC Environmental Justice Index (EJI), which aggregates demographic indicators into spatially bounded units. For example, EP_MINRTY refers to the percentage of minority persons in a given area, calculated as the proportion of all persons who do not identify as non-Hispanic White. This variable includes African American, Hispanic/Latino, Asian, Pacific Islander, Native American, and other non-White populations. EP_AFAM specifically represents the percentage of African American residents. These variables are mapped per spatial unit using EJI methodology, which corresponds to census block group aggregations.

In the initial analysis phase, correlation analyses were conducted to preliminarily explore how environmental factors and housing prices together impact levels of racial segregation within communities. However, as correlation analysis only captures pairwise relationships, it does not account for cumulative or interactive effects among variables. To address this limitation, we used regression and machine learning methods in later phases to capture more complex relationships. Subsequently, multiple Ordinary Least Squares (OLS) regression models were employed to further investigate the influence of environmental and housing price variables on the proportion of various minority populations. While the OLS model is useful for estimating linear associations, it still does not capture interaction effects or non-linear dependencies. To examine potential interactions among environmental and economic factors—as suggested by the study’s framing—we employed machine learning models capable of modeling such complexity. Specifically, Gradient-Boosted Decision Trees and Deep Neural Networks were used, which inherently consider interactions among variables during training. Additionally, deep neural network models were utilized to probe potential causal relationships between housing prices and minority segregation. Although the econometric and machine learning models are applied in sequence, rather than through a true hybrid framework, they serve complementary purposes; econometric models provide baseline interpretability, while machine learning models enable non-linear prediction and interaction discovery. This sequential yet synergistic approach allows us to confirm associations and identify high-performing predictive structures.

To deepen the analysis of these complex relationships, the study also incorporated machine learning and deep learning algorithms. To prevent overfitting and ensure good generalization of the models, random training and testing datasets were automatically generated. Model parameters were systematically optimized through cross-validation and grid search techniques, exploring various optimization algorithms to identify the most effective model configurations. Moreover, the dynamic relationships between different variables and the interpretation of model features were thoroughly examined using SHAP (SHapley Additive exPlanations). This model accurately predicts the concentration levels of minority populations and the degree of residential segregation affected by environmental factors and housing prices, providing scientific evidence and concrete recommendations for promoting environmental justice in urban planning and design.

2.1. Study Site

The city of Las Vegas, due to its unique business ecosystem and economic environment, along with rapid growth in housing market supply and commercial economy, is considered an ideal site for segregation research [61]. For more detailed analysis, the city has been divided into a grid system comprising 200 m × 200 m spatial units. Utilizing the spatial joining capabilities of Geographic Information Systems (GIS), data have been aggregated within these units, totaling 3832 spatial units. This structured approach significantly facilitates detailed spatial analysis across different urban areas (Figure 2).

2.2. Framework and Data

2.2.1. Dependent Variable

Previous research has discussed indicators that measure the degree of social isolation in different dimensions, such as homogeneity, exposure, concentration, and clustering [62]. For this research, the ‘Percentage of minority persons’ from the CDC Environmental Justice Index (EJI) dataset based on the 2022 edition is used as a metric for concentration (Figure 3).

We used the natural breaks (Jenks) method in GIS pro to display the results. These data are derived from the Census Bureau’s American Community Survey (ACS). Descriptive statistics and information for all variables can be seen in Appendix A:

Table A1. Descriptive statistics for all variables.

Var. Name	Description	Data Source	Mean	S.D.	Max	Min	Count
Dependent variable
EP_HISP	Percentage of Hispanic or Latino persons estimate	2016–2020 ACS	31.64	20.82	89.70	2.20	3832
EP_AFAM	Percentage of Black/African American (not Hispanic or Latino) persons estimate	2016–2020 ACS	11.09	8.50	52.30	0.80	3832
EP_ASIAN	Percentage of Asian (not Hispanic or Latino) persons estimate	2016–2020 ACS	6.40	3.91	29.40	0.00	3832
EP AIAN	Percentage of American Indian or Alaska Native (not Hispanic or Latino) persons estimate	2016–2020 ACS	0.47	0.79	5.40	0.00	3832
EP_NHPI	Percentage of Native Hawaiian or Other Pacific Islander (not Hispanic or Latino) persons estimate	2016–2020 ACS	0.65	1.23	7.80	0.00	3832
EP_MINRTY	Percentage minority (Hispanic or Latino) of any race; Black and African American (Not Hispanic or Latino); American Indian and Alaska Native, Not Hispanic or Latino; Asian (Not Hispanic or Latino); Native Hawaiian and Other Pacific Islander (Not Hispanic or Latino); Two or More Races (Not Hispanic or Latino); Other Races (Not Hispanicor Latino) estimate	2016–2020 ACS	53.31	22.28	97.40	7.50	3832
EP_OTHERRACE	All other populations excluding Hispanics	2016–2020 ACS	68.36	20.82	97.80	10.30	3832
Independent varable
BE
POI Count	Counts points of interest in specified areas.	Openstreet Map	0.20	0.97	25.00	0.97	3832
Zillow Houseing Price	Average house price in a specific area.	Zillow	403,155.17	135,839.99	815,000.00	50,000.00	3832
tree canopy	Percentage of area covered by tree canopy, reflecting urban forestry levels.	Nevada Division of Forestry	138.63	907.68	27,357.62	907.68	3832
air (PM2.5)	Measure of air quality PM2.5.	Air Quality-Amber/Ambee Environmental APIs	37,741.50	207.94	38,521.00	37,383.00	3832
Tree Count	Number of trees in an area.	Earth Defines Tree Locations dataset	109.76	51.34	346.00	1.00	3832
Boundary (fence + railing)	Proportion of an area marked by physical boundaries like fences and railings.	Google Street View	15.12	9.68	87.00	20.00	3832
LST	Land Surface Temperature.	EPA weather station	83.30	1.78	88.00	77.00	3832
bad site	Proximity to high-risk environmental sites, quantifying land coverage within one mile of these sites.	CDC/ATSDR Environmental Justice Index	15.41	34.90	200.00	0.00	3832

.

2.2.2. Independent Variable

Environmental variables in this study encompass six factors. The annual average PM2.5 concentration serves as an indicator for air quality assessment, derived via Bayesian Kriging interpolation, which is a method found to surpass traditional geospatial interpolation techniques in accuracy [63]. This approach utilizes data from Air Quality-Amber, providing a spatial representation of air pollution levels with annual mean PM2.5 data from the years 2018–2023 sourced from Ambee Environmental APIs. Finally, zonal statistics and the spatial join tool in GIS pro were used to summarize these data to spatial units.

In addition, the research compiles and quantifies various pollution sources, drawing on CDC environmental justice statistics. It evaluates the coverage proportion within a one-mile radius around critical environmental sites, including those listed in the EPA Toxic Release Inventory; EPA treatment, storage, and disposal facilities; EPA Risk Management Plan sites; and coal and lead mines. Such an evaluation is instrumental for mapping the spread of environmental risks throughout urban terrains, offering a granular view of environmental justice concerns.

Information regarding the distribution of individual trees in Las Vegas is collected from the Earth Defines Tree Locations dataset. These data are then aggregated to the spatial unit level using spatial join techniques, providing insights into the distribution of urban green spaces. Canopy coverage data in Las Vegas as a replenishment, obtained from the Nevada Division of Forestry, reflects the extent of tree coverage and its role in mitigating urban heat and enhancing air quality across different communities.

Google Street View also serves as an effective data source for assessing urban environments [37]. The proportion of fences and railings visible within Google Street View images is calculated to evaluate physical barriers in urban areas. Coordinates are determined by dividing streets, with each point serving as a vantage point for capturing Street View images, resulting in a dataset of 103,170 Street View Images (SVIs).

To preprocess and extract features from these images, the study employed the Pyramid Scene Parsing Network (PSPNet), which is a deep convolutional neural network model, to do the segmentation. This framework has been widely used in urban research. The PSPNet model was pre-trained on the ADE20K database, a comprehensive dataset of SVI data collected from 50 cities. This preprocessing method yielded ratios of different physical elements in streetscapes relative to the whole image. This dataset and the derived features serve as an environmental measure of physical distancing.

Simultaneously, land surface temperature (LST) offers a unique perspective on environmental resource inequality. Data on LST are sourced from the U.S. Environmental Protection Agency, providing essential insights into the urban heat island effect and climate adaptation measures. Zonal statistics and the spatial join tool in GIS pro were used to summarize these data to grid cell units.

Housing price data for this study were obtained from Zillow, a leading online real estate marketplace in the United States. Using the crawler tool, we downloaded the data within 3 years according to the zip code of the whole city of Las Vegas and calculated the average price of each space unit as the housing price index (Figure 4). Also, the global spatial autocorrelation test indicates that the housing price is not randomly distributed, rather, it likely reflects spatial heterogeneity arising from underlying economic conditions that differ across regions. This variable exhibits a global Moran’s I of 0.097, with a p-value < 0.001, showing a significant positive correlation. In other words, areas with high (or low) prices tend to be adjacent to areas with similarly high (or low) prices, forming spatial clusters. Recognizing this aggregation is crucial, as it allows for a deeper explanation in the subsequent interpretation of results.

3. Results

3.1. Normality Test

For the data preprocessing, this study implemented a method based on the 1.5 interquartile range (IQR) to remove outliers from the price variable. The normality of all variables was generally assessed using the Shapiro–Wilk test and the Kolmogorov–Smirnov (K-S) test. Given its greater sensitivity for small samples, the Shapiro–Wilk test was prioritized for interpretation. Since many variables exhibited skewness and non-normality (p < 0.05), a symmetric log-transformation was selectively applied to highly skewed variables, such as housing price and tree count, based on diagnostic plots and theoretical justification. Log-transforming housing prices is particularly appropriate given the right-skewed distribution of property values in urban markets. To further prepare the data for machine learning algorithms and mitigate the influence of extreme values, all variables were standardized using z-scores. While standardization adjusts for scale and central tendency, it does not correct skewness, reinforcing the need for prior transformation. Consequently, a Spearman correlation matrix was employed for preliminary assessment of pairwise relationships, given the predominantly non-normal distribution of the features.

3.2. Correlations

In this statistical analysis, the correlations between the minority population proportion (EP_MINRTY) and thirteen other variables were examined using Spearman’s rank correlation co-efficient, appropriate for non-normally distributed data.

The correlation results (Appendix A:

Table A2. Spearman correlation.

	1	2	3	4	5	6	7	8	9	10	11	12	13	14
Price (1)	1
EP_MINRTY (2)	−0.531 **	1
EP_AFAM (3)	−0.147 **	0.426 **	1
EP_HISP (4)	−0.509 **	0.900 **	0.212 **	1
EP_ASIAN (5)	0.278 **	−0.268 **	−0.005	−0.383 **	1
EP_AIAN (6)	−0.005	−0.068 **	−0.106 **	−0.071 **	−0.009	1
EP_NHPI (7)	−0.050 **	0.196 **	0.022	0.195 **	−0.057 **	−0.034 *	1
air (8)	−0.331 **	0.518 **	−0.056 **	0.472 **	−0.012	−0.082 **	−0.030	1
Tree_Count (9)	0.282 **	−0.407 **	−0.181 **	−0.411 **	0.298 **	−0.057 **	−0.093 **	−0.087 **	1
boundary (10)	−0.245 **	0.391 **	0.199 **	0.394 **	−0.214 **	0.026	−0.014	0.171 **	−0.246 **	1
LST (11)	−0.414 **	0.609 **	0.155 **	0.612 **	−0.295 **	0.030	0.114 **	0.386 **	−0.419 **	0.277 **	1
bad_site (12)	−0.136 **	0.343 **	0.108 **	0.268 **	−0.035 *	−0.018	−0.016	0.401 **	−0.147 **	0.157 **	0.304 **	1
POI_Count (13)	−0.125 **	0.149 **	0.081 **	0.115 **	−0.019	−0.012	−0.016	0.136 **	−0.175 **	0.015	0.141 **	0.113 **	1
treecanopy (14)	−0.032	0.048 **	0.018	0.058 **	−0.000	0.006	0.030	0.014	−0.004	0.014	0.045 **	−0.033 *	−0.009	1

* p < 0.05 ** p < 0.01.

) indicate that neighborhoods with higher minority populations are generally associated with lower housing prices, poorer air quality (higher PM2.5), and elevated land surface temperatures. These patterns suggest persistent spatial inequalities in environmental and economic conditions that disproportionately affect minority communities. Tree count and the proportion of Asian residents (EP_ASIAN) also exhibit positive correlations with EP_MINRTY, indicating that some minority-concentrated neighborhoods may benefit from better urban vegetation and reflect higher Asian community presence. All relationships identified were statistically significant at the 0.01 level. The negative correlation between house prices and EP_HISP suggests that house prices tend to be lower in areas with a higher percentage of Hispanic residents, reflecting broader patterns of socioeconomic stratification and spatial exclusion. In addition, EP_MINRTY is strongly correlated with both house prices and EP_HISP, with coefficients exceeding 0.5. Of these, the correlation between EP_MINRTY and EP_HISP is very high (close to 1), which underscores the demographic reality that Hispanic or Latino populations make up a significant portion of the total minority population in Las Vegas. While these pairwise correlations are insightful, they do not capture interaction effects or control for confounding variables. To address this issue, we used more comprehensive modeling techniques such as the multiple regression and machine learning methods applied below.

3.3. Ordinary Least Squares

Given the presence of skewness and negative values in the dataset, a symmetric log transformation was selectively applied to variables such as housing price and tree count, based on diagnostic assessments and theoretical justification. This transformation aimed to mitigate heteroscedasticity, reduce the influence of outliers, and stabilize variance across the data prior to conducting OLS and machine learning model analyses. Subsequently, OLS regression (Equation (1)) was employed to investigate variations in the concentration of different ethnic groups. This methodological approach provides a more nuanced exploration of ethnic concentration patterns, potentially revealing disparities that simpler models might overlook.

$$\begin{gathered} ethnicgrou{p}_i={\beta }_0+{\beta }_1{\mathrm{P}\mathrm{r}\mathrm{i}\mathrm{c}\mathrm{e}}_i+{\beta }_2{\mathrm{b}\mathrm{a}\mathrm{d}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{e}}_i+{\beta }_3{ \mathrm{treecanopy}}_i+{\beta }_4{\mathrm{P}\mathrm{O}\mathrm{I}\mathrm{c}\mathrm{o}\mathrm{u}\mathrm{n}\mathrm{t}}_i \\ +{\beta }_5{\mathrm{a}\mathrm{i}\mathrm{r}}_i+{\beta }_6{\mathrm{T}\mathrm{r}\mathrm{e}\mathrm{e}\mathrm{c}\mathrm{o}\mathrm{u}\mathrm{n}\mathrm{t}}_i+{\beta }_7{\mathrm{b}\mathrm{o}\mathrm{u}\mathrm{n}\mathrm{d}\mathrm{a}\mathrm{r}\mathrm{y}}_i+{\beta }_8{\mathrm{L}\mathrm{S}\mathrm{T}}_i+{\epsilon }_i \end{gathered}$$

(1)

In this equation, $ethnicgrou{p}_i$ represents six ethnicity indicators: EP_HISP, EP_AFAM, EP_ASIAN, EP_ALAN, EP_NHPI and EP_MINRTY. The detailed $\beta$ coefficients results can be found in the regression results table,

Table 1. Ordinary Least Squares.

Variables	EP_HISP			EP_AFAM			EP_ASIAN			EP_ALAN			EP_NHPI			EP_MINRTY
	Coef.	S.E	t	Coef	S.E.	t	Coef.	S.E.	t	Coef.	S.E.	t	Coef.	S.E.	t	Coef.	S.E.	t
constant	−0.032 **	0.007	−4.49	−0.028 **	−0.028	−3.023	−0.011	0.01	−1.101	−0.078 **	0.009	−8.236	−0.094 **	0.009	−10.306	−1423.357 **	53.838	−26.438
Price	−0.213 **	0.012	−17.502	−0.098 **	−0.098	−6.132	0.1788 **	0.017	10.408	−0.023	0.016	−1.42	−0.054 **	0.016	−3.414	0.000 **	0.000	−11.674
bad_site	−0.055 **	0.014	−4.049	0.226 **	0.226	12.682	−0.0888 **	0.019	−4.651	0.032	0.018	1.776	0.03	0.018	1.715	0.060 **	0.008	7.444
treecanopy	0.009	0.028	0.332	0.001	0.001	0.014	0.048	0.039	1.249	0.004	0.037	0.11	0.064	0.036	1.776	0.000	0.000	−0.374
POI_Count	−0.046 *	0.019	−2.496	0.098 **	0.098	4.033	0.1058 **	0.026	4.05	0.016	0.025	0.645	−0.058 *	0.024	−2.423	−0.300	0.279	−1.075
air	0.379 **	0.013	29.025	−0.229 **	−0.229	−13.399	0.1528 **	0.018	8.336	−0.136 **	0.017	−7.813	−0.098 **	0.017	−5.795	0.032	0.001	21.636
Tree_Count	−0.177 **	0.012	−14.568	−0.111 **	−0.111	−6.97	0.1868 **	0.017	10.943	0.015	0.016	0.95	−0.019	0.016	−1.208	−0.048 **	0.006	−8.056
boundary	0.14 **	0.011	12.154	0.117 **	0.117	7.796	−0.0988 **	0.016	−6.103	−0.018	0.015	−1.201	−0.074 **	0.015	−4.956	0.257 **	0.029	8.964
LST	0.302 **	0.013	22.986	0.077 **	0.077	4.467	−0.1538 **	0.018	−8.337	0.005	0.018	0.276	0.094 **	0.017	5.549	3.218 **	0.190	16.942
R 2	−0.032			0.143			0.173			0.019			0.026			0.594
Adjusted R 2	−0.213			0.141			0.171			0.017			0.024			0.591

* p < 0.05; ** p < 0.01.

.

This study uses the Ordinary Least Squares (OLS) multivariate method to explore the relationship between the concentration of different racial groups and various environmental and socioeconomic factors (Table 1). The results indicate that factors such as housing prices, air quality, the number of points of interest (POIs), and land surface temperature (LST) play crucial roles in shaping patterns of residential segregation. Specifically, housing prices are negatively associated with the concentration of the EP_HISP group, with a coefficient of −0.213 (t = −17.502, p < 0.0001), indicating that in areas with higher housing prices, the proportion of Hispanic residents tends to decrease. Conversely, housing prices are positively associated with EP_ASIAN, with a coefficient of 0.178 (t = 10.408, p < 0.0001), suggesting that higher-cost areas are more likely to have a greater concentration of Asian residents.

In terms of environmental quality, air quality has a significant positive impact on the EP_HISP group, with each unit increase in air quality leading to a 0.379 unit increase in ethnic concentration (t = 29.025, p < 0.0001). However, for the EP_AFAM group, an improvement in air quality results in a decrease in concentration, with a coefficient of −0.229 (t = −13.399, p < 0.0001), highlighting divergent environmental exposure patterns across racial groups. Additionally, poor site conditions (bad site) have a positive effect on the concentration for EP_AFAM, with a coefficient of 0.226 (t = 12.682, p < 0.0001), suggesting that African American populations are more likely to reside in environmentally degraded areas. For the EP_NHPI group, a higher number of trees and more physical boundaries (e.g., fences, railings) are negatively associated with concentration. These variables also show similar or stronger negative associations with EP_HISP and EP_AFAM (Table 1).

The overall explanatory power of the model shows significant differences among different racial groups, with the model fit being highest for the Hispanic group (adjusted R² of 59.1%) and lowest for the Native Hawaiian or Other Pacific Islander group (adjusted R² of 2.4%). These differences reflect the varying degrees to which modeled variables explain racial group concentrations. The very low adjusted R² values for groups such as EP_NHPI and EP_AIAN may be related to small population sizes and unobserved contextual factors.

Overall, the regression analysis clearly demonstrates that housing prices and environmental quality play a crucial role in shaping the patterns of concentration among different racial groups. These findings underscore the need for urban planning and policy strategies that actively address environmental injustice and economic exclusion to foster more equitable and inclusive community development.

3.4. Causal Relationship Estimation Processing Effect

In this study, to measure the potential causal relationship and estimate the treatment effect, we employed a neural network model called TARNet (Treatment-Agnostic Representation Network) to estimate the causal effect of price changes on the economic status of minority groups. The TARNet model learns two distinct outcome paths through a shared representation layer (phi), with one path corresponding to the treatment group and the other to the control group [64]. This enables the model to learn different treatment effects based on the same input. The model uses regularization to prevent overfitting and utilizes Mean Squared Error (MSE) as the loss function. Through this model, we calculated the Average Treatment Effect (ATE) and Conditional Average Treatment Effect (CATE), both of which were −0.8348584, indicating a significant average difference in the EP_MINRTY outcome variable between the treatment group with prices above the median and the control group with prices below the median. This suggests that an increase in price had an average negative impact on the outcome variable.

The detailed visualization of the model’s multilayer structure using TensorFlow Keras utils is presented in Figure 5. The TARNet architecture begins with an input layer accepting 12 features, followed by 2 shared representation dense layers (200 neurons each) that form the representation network. This network then branches into two parallel pathways: one for the control group (left branch) and one for the treatment group (right branch). Each pathway contains a hidden layer (100 neurons) with ELU activation and L2 regularization, followed by a single-neuron output layer with linear activation. This dual-path structure enables the model to simultaneously predict outcomes for both treatment conditions, facilitating the calculation of Average Treatment Effect (ATE) and Conditional Average Treatment Effect (CATE). The model contains 83,202 trainable parameters in total, optimized using the Adam optimizer, with mean squared error as the loss function (

Table 2. Causal effects of housing prices on minority population concentration.

Metric	Value
ATE	−0.8348584
CATE	−0.8348584
Standard Deviation of CATE	0.74378604
Total Model Parameters	83,202

).

The model’s further analysis reveals that the standard deviation of the Conditional Average Treatment Effect (CATE) is 0.74378604. This suggests that while the overall trend indicates a significant average effect of housing price changes on minority groups, the effect size might vary significantly among individuals or specific subgroups. This variation indicates that the impact of housing price changes on the proportion of minority groups may exhibit spatial heterogeneity across different areas or socioeconomic backgrounds. Consequently, future research should focus on a more detailed analysis of the effects of housing price changes, considering the specific circumstances of the regions and socioeconomic factors.

The above results not only emphasize the need to consider the broad socioeconomic impacts when formulating real estate policies but also highlight the importance of accounting for potential individual and group differences when evaluating policy outcomes. In this way, policymakers can more accurately understand and predict the social consequences of economic decisions, leading to the formulation of more equitable and effective strategies.

3.5. Prediction Models

3.5.1. Standard Model Performance

Following the optimization of hyperparameters via random search and subsequent validation on the validation set, the predictive performance of the model was assessed using the test set. For hyperparameter optimization, the study employed RandomizedSearchCV with 3-fold cross-validation and 10 iterations across all models. The specific configurations encompass a wide range of parameters to ensure comprehensive model exploration. The Decision Tree model’s hyperparameter ranges include max_depth = [None,10, 20, 30]; min_samples_split = [2, 10, 20]; and min_samples_leaf = [1, 5, 10], aiming to explore the balance between model complexity and generalization capability. For Random Forest, we examined n_estimators = [10, 50, 100, 200]; max_features = [‘auto’, ’sqrt’]; max_depth = [None,10, 20, 30, 40, 50]; min_samples_split = [2, 5, 10]; and min_samples_leaf = [1, 2, 4] to fully leverage ensemble advantages while preventing overfitting. The MLP Regressor exploration included different network architectures (hidden_layer_sizes = [(50,), (100,), (50, 50), (100, 50)]); activation functions ([‘tanh’, ’relu’]); optimizers ([‘sgd’, ’adam’]); regularization strengths (alpha = [0.0001, 0.001, 0.01]); and learning rates ([0.001, 0.01, 0.1]). SVR (Support Vector Regression) configurations include kernel = [‘linear’, ’poly’, ’rbf’, ’sigmoid’]; regularization parameter = [0.1, 1, 10, 100]; and gamma = [‘scale’, ’auto’] to capture potential non-linear segregation patterns. For the LGBM Regressor, tuned parameters include num_leaves = [31, 127]; max_depth = [10, 20, 30]; learning_rate = [0.01, 0.05, 0.1]; and n_estimators = [100, 200, 500].

Model performance was evaluated through metrics such as MSE, Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Determination (R²). Additionally, the model’s predictive accuracy was visually assessed by plotting scatter diagrams comparing actual versus predicted housing prices.

$$MSE={\displaystyle \frac{1}{n}}\sum _{i=1}^n {\left(y_i-{\widehat{y}}_i\right)}^2$$

(2)

$$RMSE=\sqrt{MSE}$$

(3)

$$MAE={\displaystyle \frac{1}{n}}\sum _{i=1}^n \left|y_i-{\widehat{y}}_i\right|$$

(4)

$$MAPE={\displaystyle \frac{100\%}{n}}\sum _{i=1}^n\left|{\displaystyle \frac{y_i-{\widehat{y}}_i}{y_i}}\right|$$

(5)

$$R^2=1-{\displaystyle \frac{{\sum }_{i=1 }^n{\left(y_i-{\widehat{y}}_i\right)}^2}{{\sum }_{i=1}^n {\left(y_i-\overline{y}\right)}^2}}$$

(6)

The dataset was split, allocating 80% for model training and the remaining 20% to form a provisional dataset for further division. This provisional dataset was evenly split again: half serving as the validation set (X_val, y_val) for model tuning and verification and the other half as the test set (X_test, y_test) for evaluating the model’s final performance. This bifurcation ensures that the test set’s independence is maintained, enabling a reliable estimation of the model’s performance.

The study’s comparative analysis tested five machine learning models: Decision Tree Regressor, Random Forest Regressor, MLP Regressor, SVR, and LGBM Regressor. After hyperparameter optimization using a random search on the validation set and final evaluation on the test set, the LGBM Regressor model demonstrated superior performance, achieving an R² of 0.798. This indicates a high explanatory power concerning housing price variations. The model also performed well on other performance metrics, signifying its effectiveness in capturing underlying data patterns (

Table 3. Comparison of machine learning results.

	R2	MSE	RMSE	MAE	MAPE
Decision Tree Regressor	0.722	0.114	0.338	0.232	1.336
Random Forest Regressor	0.775	0.092	0.304	0.230	1.450
MLP Regressor	0.712	0.117	0.343	0.259	1.665
SVR	0.675	0.133	0.365	0.267	1.667
LGBM Regressor	0.798	0.082	0.288	0.203	1.287
Gaussian Process Regressor	−0.001	0.411	0.641	0.563	1.0

).

Based on the metrics, LGBM Regressor and Random Forest Regressor are highly effective models for this dataset, providing a strong balance between complexity and prediction accuracy. Decision trees offer a simpler, though less accurate alternative. MLP, SVR, and particularly Gaussian Process do not perform as well in this context, suggesting that tree-based methods (LGBM Regressor) might be more suitable for the structure of the data and this prediction task (Figure 6).

3.5.2. Deep Learning Model

Deep learning models, particularly those utilizing ReLU activation functions, excel at capturing non-linear relationships between variables, which are common in social and economic data. This capability is crucial for understanding the multifaceted and interdependent factors leading to racial concentration [65]. In this study, we attempt to compare different optimization algorithms to assess their predictive performance. For regularization techniques, this study applied a dropout rate of 0.2 after each dense layer. Batch normalization was implemented after each fully connected layer, helping to stabilize the training process and accelerate convergence. The deep learning predictive model was constructed using the TensorFlow and Keras deep learning frameworks, based on a fully connected feedforward neural network. The architecture of the model comprises an input layer, several hidden layers, and an output layer. Each hidden layer employs the ReLU activation function, and some layers are equipped with Batch Normalization and Dropout techniques. Batch Normalization is applied to standardize the activations of layers to expedite training and enhance stability, while Dropout randomly sets the output of some neurons to zero during training to reduce overfitting. Specifically, the model begins with a fully connected layer consisting of 64 neurons, followed by a batch normalization layer and a 20% Dropout layer. This is succeeded by a 32-neuron fully connected layer, another application of batch normalization and Dropout, and finally, a single-neuron output layer that yields continuous predictive values for prices. The network comprises multiple layers, each defined by linear transformations and non-linear activation functions. The mathematical description of each dense layer in the network is as follows.

For each fully connected Layer (Dense Layer):

$$z^{\left[l\right]}=W^{\left[l\right]}{a}^{[l-1]}+b^{\left[l\right]}$$

(7)

$$a^{\left[l\right]}=g^{\left[l\right]}\left(z^{\left[l\right]}\right)$$

(8)

In Equation (7), $z^{\left[l\right]}$ represents the linear output of layer 1, $W^{\left[l\right]}$ is the weight matrix, $a^{[l-1]}$ is the activation output from the previous layer, and $b^{\left[l\right]}$ is the bias vector. In Equation (8), $a^{\left[l\right]}$ denotes the activation output of layer 1 and $g^{\left[l\right]}$ is the activation function.

The training of the model employed optimization algorithms such as Stochastic Gradient Descent (SGD), Adam, RMSprop, and Adagrad. Each algorithm was configured with distinct learning rates and hyperparameters to assess their impact on model performance. The Mean Squared Error (MSE) served as the loss function to minimize the difference between predicted and actual values during training. SGD, the most fundamental optimization method, updates parameters as follows:

$$\mathrm{SGD}: {\theta }_{t+1}={\theta }_t-\eta \cdot {\nabla }_{\theta }J\left(\theta \right)$$

(9)

For the SGD update rule (Equation (9))$, {\theta }_t$ represents the parameter value at time t, $\eta$ is the learning rate, and ${\nabla }_{\theta }J\left(\theta \right)$ is the gradient of the loss function J with respect to parameters $\theta$.

Adam (Adaptive Moment Estimation) combines the advantages of momentum and RMSprop. Its parameter update rule is

$$m_t={\beta }_1\cdot m_{t-1}+\left(1-{\beta }_1\right)\cdot {\nabla }_{\theta }J\left(\theta \right)$$

(10)

$$v_t={\beta }_2\cdot v_{t-1}+\left(1-{\beta }_2\right)\cdot {\left({\nabla }_{\theta }J\left(\theta \right)\right)}^2$$

(11)

$${\widehat{m}}_t={\displaystyle \frac{m_t}{1-{\beta }_1^t}}$$

(12)

$${\widehat{v}}_t={\displaystyle \frac{v_t}{1-{\beta }_2^t}}$$

(13)

$${\theta }_{t+1}={\theta }_t-{\displaystyle \frac{\eta }{\sqrt{{\widehat{v}}_t}+ϵ}}\cdot {\widehat{m}}_t$$

(14)

In Equation (10), $m_t$ is the first moment estimate (momentum) at time t and ${\beta }_1$ is the exponential decay rate for the first moment. In Equation (11), $v_t$ is the second moment estimate at time t and ${\beta }_2$ is the exponential decay rate for the second moment. In Equation (12), ${\widehat{m}}_t$ is the bias-corrected first moment estimate. Regarding Equation (13), ${\widehat{v}}_t$ is the bias-corrected second moment estimate. In Equation (14), $ϵ$ is a small constant added for numerical stability.

RMSprop is an improvement on Adagrad for better performance in non-convex settings. Its update rule is

$$v_t=\beta ·v_{t-1}+\left(1-\beta \right)·{\left({\nabla }_{\theta }J\left(\theta \right)\right)}^2$$

(15)

$${\theta }_{t+1}={\theta }_t-{\displaystyle \frac{\eta }{\sqrt{v_t}+ϵ}}\cdot {\nabla }_{\theta }J\left(\theta \right)$$

(16)

In Equation (15), $v_t$ is the moving average of squared gradients and $\beta$ is the decay rate. In Equation (16), parameters are updated using the normalized gradient.

The model was trained on the training set and validated on a validation set, constituting 20% of the data, to adjust model parameters and prevent overfitting (Figure 7). Monitoring the change in validation loss during training, the model’s performance was then evaluated on the test set upon completion of training.

In comparing the performance of deep learning frameworks with traditional machine learning algorithms in terms of interpretability, our results indicate that while deep learning models generally maintain a stable R² value above 0.7 (

Table 4. Comparison of deep learning results.

Deep Learning Results
	R2	MSE	RMSE	MAE	MAPE
SGD	0.717	0.113	0.336	0.252	1.510
Adam	0.735	0.106	0.325	0.244	1.500
RMSprop	0.729	0.108	0.328	0.242	1.789
Adagrad	0.719	0.249	0.499	0.364	1.720

), they do not significantly outperform the gradient-boosted LGBM Regressor model when data are scarce. This phenomenon could be attributed to several factors. Deep learning models typically require a substantial amount of data to support their complex feature-learning capabilities; with limited data, they may not be sufficiently trained, leading to suboptimal performance. Furthermore, due to their high model complexity, deep learning models are more prone to overfitting, especially with fewer samples. The choice of optimization algorithms also significantly impacts performance, as different optimizers, such as SGD, Adam, RMSprop, and Adagrad, exhibit varied outcomes on performance metrics (Figure 7). These elements collectively suggest that traditional machine learning models with simpler architectures and inherent regularization properties, like the LGBM Regressor, may have an advantage in data-constrained environments. This finding underscores the importance of matching model complexity with data scale and considering data limitations when selecting appropriate models and algorithms.

In order to explain the model mechanism study using the best estimator from the top-performing LGBM Regressor, a SHAP interpreter was established to facilitate the summary beeswarm plot of features from SHAP importance analysis based on the Model. This plot graphically depicts the relative impact of each feature across the entire dataset, providing a comprehensive understanding of their influences. The beeswarm plot (Figure 8) offers an intuitive visual representation of feature importance, helping to pinpoint key factors contributing to the model’s predictions and supporting insightful, data-driven decision-making.

The results indicate a strong negative correlation between the proportion of minorities and housing prices, which aligns with previous discussions on potential residential segregation among minorities. At the same time, long-term macro-environmental exposure issues, such as air pollution and surface temperature, exhibit a significant positive relationship with the concentration of minorities. Interestingly, the number of points of interest (POIs) does not have substantial explanatory power. Overall, the selected variables demonstrate strong explanatory power in predicting the concentration of minorities, consistent with our previous conclusions that minority-concentrated areas exhibit distinct environmental characteristics.

To further investigate the impact of critical features on overall prediction results and their interaction effects with other features, a SHAP partial dependence plot was constructed, as shown in the figure. The plot reveals a notable interaction effect between different features.

In areas with higher housing prices, the surface temperature tends to be lower, possibly due to better infrastructure or greener environments. Interestingly, in Las Vegas, there is a negative correlation between the number of trees and the SHAP values of housing prices. As the number of trees increases, the impact on prices becomes less positive or even negative, suggesting that areas with many trees may not align with higher real estate prices. Areas with a higher number of trees may buffer or amplify the effects of temperature changes, while limited planting space restricts the potential for vegetation-mediated cooling [66]. Higher canopy cover is associated with lower SHAP values for air quality, indicating that areas with higher canopy cover tend to have better air quality (Figure 9). However, these areas are also associated with a significant negative correlation with minority concentration, suggesting a potential increased risk of environmental pollution exposure. Additionally, there is a positive relationship between bad sites and air quality, whereas the correlation between boundary and bad sites is negative.

4. Discussion

This study employs a finer spatial resolution at the street level, allowing for more detailed and localized insight into segregation dynamics, while much of the previous literature has focused on analyzing segregation at the census tract or neighborhood level.

Secondly, rather than relying solely on static socioeconomic indicators, this study integrates real-time, multidimensional data sources, including Google Street View imagery, POI distributions, house prices, and environmental quality indicators such as PM2.5 and surface temperature. Apart from that, this study integrates economic data and environmental variables and combines traditional econometric models with advanced machine learning techniques, illuminating how urban system elements can exacerbate or mitigate segregation. In addition, the study introduces a deep learning-based causal inference model (TARNet) to explore the correlation and potential causal effects between house price dynamics and minority distribution. This perspective is essential for policymakers, underscoring the need for both environmental and socioeconomic considerations in real estate policy formulation to ensure equitable community outcomes.

Furthermore, our analysis sheds light on the intricate interplay between environmental features and economic forces within urban settings. The significant role of these factors in shaping residential patterns emphasizes the critical need for urban planning and policy that prioritizes environmental justice and sustainability. Additionally, not only is the impact of streetscape attributes on segregation highlighted but the importance of understanding these dynamics to create more informed and just urban policies is emphasized.

4.1. Multiple Segregations at the Physical and Economic Levels

Over the past two decades, residential segregation has remained a central focus for one of the most pivotal academic institutions in social science. Much like the suburbanization movement of the 1950s, Las Vegas has experienced a parallel process of segregation in its physical environmental attributes, borne from environmental construction and commercial development [35]. The income class composition of gated and non-gated block groups in Las Vegas is notable. In Las Vegas, the residences of the upper income class constitute a high proportion gated housing structure [35].

This study delves into the issue of residential segregation by examining the influence of micro-level streetscape attributes (such as boundaries, tree canopy, and trees) along with macro-scale environmental exposures (air quality, land surface temperature) on the distribution of minority populations. By doing so, it extends the traditional focus beyond green spaces to include them as indicators of resource inequality. Consistent with findings from traditional research, there is a significant correlation between minority concentration and residential segregation caused by resource inequalities. Targeted interventions in future community development, such as enhancing urban greenery or improving air quality, could significantly improve conditions in minority-dense communities. Simultaneously, physical architectural features can also serve as a measure of segregation.

Furthermore, deep learning models in this study have uncovered potential causal links between housing prices and minority concentrations, revealing notable heterogeneity in these relationships. This variability underscores the importance of how changes in housing prices differentially impact minority populations across various regions and socioeconomic terrains. While theoretical perspectives may vary, the majority of scholars concur that residential segregation critically limits the resources and opportunities available to minority populations, as these remain closely intertwined with the racial/ethnic composition of American communities. A reconsideration of the distribution of local capital is necessary.

4.2. Ethnic Dynamics of Segregation

Unlike cities with diverse racial composition and segregation, such as Los Angeles, Las Vegas’ minorities are predominantly Latino [67]. Different variables affect the concentration elasticity of different races to different degrees. In the model predictions, house prices have a negative impact on the concentration share of Latinos, but a positive impact on Asians. From the perspective of environmental characteristics, good environments (e.g., high air quality) increase the racial concentration of Latinos, whereas poor environments and bounded parks, in turn, increase the concentration of African Americans. Thus, variations in these factors may be the result of different concentration influences between races due to cultural priorities, class and lifestyle habits, as well as types of social activities and perceptions of other people in the same place [68].

High housing prices tend to exclude certain groups, especially Hispanic groups-from more affluent and resource-rich neighborhoods, thus increasing residential segregation. In contrast, Asian (EP_ASIAN) communities appear to be more resilient to rising housing costs and may even benefit from higher property values, possibly due to differences in socioeconomic status or cultural preferences for urban living environments. Historically discriminatory practices in the Las Vegas real estate market, such as preserving property values by restricting minority access, have contributed to the declining concentration of Latinos and African Americans in high-value areas [21,22]. In addition, social and psychological factors may further discourage minority participation in affluent areas. Feelings of exclusion, limited opportunities for social integration, and a lack of belonging in shared public areas may prevent minorities from settling in affluent neighborhoods. These dynamics highlight how economic barriers and social exclusion combine to reinforce long-standing patterns of racial segregation [69]. In addition, differences in the distribution of social benefits contribute to patterns of socioeconomic segregation, as noted in the literature [70]. Environmental quality further exacerbates these differences as different ethnic groups have different attitudes, behaviors and social contexts towards the environment, leading to different environmental configurations and thus localized patterns of ethnic agglomeration [71].

4.3. Machine Learning Predict Concentration

In this study, the employment of various machine learning models enabled a comprehensive analysis of the data from multiple perspectives. The diversity of models facilitated the identification of the most effective approach based on performance metrics. Although validation and overfitting-preventative measures were implemented, some models demonstrated potential overfitting, particularly in cases where the complexity of the model outweighed the size and diversity of the dataset. Furthermore, more complex models, such as deep learning neural networks, may be more computationally intensive and less interpretable due to their intricate network layer structures compared to simpler models.

Deep models utilizing the ReLU activation function are adept at capturing non-linear relationships, and their hierarchical structure inherently grasps the interactive effects between different predictors. This is particularly valuable in exploring complex social phenomena such as residential segregation. After applying overfitting prevention techniques through controlled layers, the explanatory benefits may show marginal differences compared to LGBM Regressor models, yet these deep learning models achieve superior predictive performance with more stable error metrics and reliability in comparison to commonly used baseline machine learning models. However, the causal inference results derived from deep learning models have provided insights into the potential causal link between housing prices and concentration, offering a degree of interpretability.

Future endeavors may involve visualizing individual decision layers to elucidate the specific decision-making processes and operational characteristics of model features, thereby enhancing interpretability.

5. Limitations

While this study contributes valuable insights, it does have some potential limitations. The data utilized, although extensive, might not capture all the variables influencing residential segregation. There could be unobserved confounders, such as policy changes or historical factors, that impact the results. Although the SHAP partial dependence plot realizes the interactivity between different variables, the mediating effects of economic factors on environmental aspects—such as how affordability filters access to green amenities—should be explored further in future research.

Additionally, although the study incorporates multiple socio-environmental indicators, some commonly used demographic and socioeconomic variables were not included due to data availability at the spatial scale of analysis. These omitted variables include educational attainment, employment status, household composition, population density, and proximity to urban cores. Future work could enhance explanatory power and address potential omitted variable bias by integrating these important controls.

Focusing on the Las Vegas area provides an in-depth look at one urban environment but may limit the generalizability of the findings to other regions. Different cities may exhibit unique socioeconomic dynamics and urban designs that the current model does not account for. In Las Vegas, minority groups, particularly Latinos, tend to concentrate in certain areas, a pattern that may differ in other cities. The study examines intra-urban spatial variation within Las Vegas by comparing the distribution of racial group concentrations across different neighborhoods or grid cells. These comparisons are modeled using environmental and economic factors such as housing prices, pollution exposure, and tree coverage. While the study is limited to a single urban context and does not provide longitudinal data, the findings reveal patterns of inequality that warrant further scholarly investigation. Combining geographically weighted machine learning models with traditional spatial lag regression models may provide richer and more comparable results.

Moreover, the cross-sectional nature of the study prevents an understanding of the temporal dynamics of residential segregation. Longitudinal studies could provide insights into how the relationships between streetscape attributes and segregation evolve over time, especially considering the evolving settlement patterns of Latino populations.

Lastly, while the study includes various socioeconomic and environmental variables, the inclusion of subjective measures such as perceived neighborhood quality or resident satisfaction could enrich the analysis. Addressing these limitations in future research could offer a more nuanced understanding of residential segregation and inform more targeted policy interventions.

6. Conclusions

This study forecasts the concentration of minority groups by constructing traditional econometric models and comparing various machine learning and deep learning algorithms. Among them, LGBM demonstrated the highest predictive performance, highlighting its suitability for modeling complex, non-linear spatial patterns of segregation. Through SHAP mechanism analysis, it offers valuable insights into the contribution of model input features to concentration, highlighting the dominant role of macro-environmental exposure (air quality, land surface temperature, and green space availability) and housing prices. Finally, the study explores potential causal conclusions of housing segregation and possible spatial heterogeneity features through deep learning-based causal inference. The result suggests a negative causal effect of increasing housing prices on minority population concentration. This framework provides strong insights for future urban renewal and planning stakeholders, while also making inferences at a more granular scale, thereby supplementing prior research conducted at the census tract level. At the same time, our conclusion indicates that the differences in isolation in these areas are universally present, regardless of the indicators used, and we need more detailed policies to help mitigate the issue of isolation.