1# RECORD ee4a73efc5ad67b4400db23b09eeacbeb4812f862# TITLE: Explainable AI-based mass appraisal: Insights from machine learning applications in Korea’s residential property market3# YEAR: 2025.0 VENUE: International Journal of Strategic Property Management DOI: 10.3846/ijspm.2025.251374# AUTHORS: Woosung Kim; Insu Choi; Myounggu Lee5# PAGES: 27 CANDIDATE TABLE PAGES (1-based): [13, 14]6# ABSTRACT: Mass appraisal plays a pivotal role in real estate management, facilitating property tax assessments, mortgage evaluations, and urban planning across large geographical areas. In regions like Korea, where real estate markets are rapidly evolving, valuation models based on multiple linear regression are valued for their simplicity and interpretability but often fall short in capturing complex market dynamics. In contrast, machine learning (ML) models, while addressing non-linear relationships between property characteristics and market values and offering superior predictive performance, are often criticized for their “black-box” nature, which raises concerns over interpretability in transparency-critical domains like property tax assessments and policy planning. To address these concerns, this study investigates the application of Explainable AI (XAI) techniques in the mass appraisal of residential properties in Korea, integrating XAI methods with both multiple linear regression and random forest models. Using SHAP (SHapley Additive exPlanations) and PFI (Permutation Feature Importance) values, the study analyzes feature importance and predictive contributions, offering insights into the factors driving property valuations. Additionally, a temporal analysis was conducted by segmenting the data into time intervals to examine how feature importance and predictive contributions evolve over time. By combining high predictive performance with transparent and interpretable insights, the findings demonstrate that XAI can enhance the usability of both traditional and advanced automated valuation models (AVMs) for real-world decision-making in the Korean real estate sector.789===== PAGE 1 (score 9, numeric-paren 0, table-words 0, var-words 9) =====1011This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), 12which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.13Copyright © 2025 The Author(s). Published by Vilnius Gediminas Technical University14ISSN: 1648-715X / eISSN: 1648-917915INTERNATIONAL JOURNAL 16OF STRATEGIC PROPERTY 17MANAGEMENT18EXPLAINABLE AI-BASED MASS APPRAISAL: INSIGHTS FROM MACHINE 19LEARNING APPLICATIONS IN KOREA’S RESIDENTIAL PROPERTY MARKET20Woo-sung KIM 1, Insu CHOI 2, Myounggu LEE 1*211 Department of Business Administration, Konkuk University, Seoul, South Korea 222 Division of Business Administration, Dongduk Women’s University, Seoul, South Korea23Article History: 24Abstract. Mass appraisal plays a pivotal role in real estate management, facilitating property tax assessments, 25mortgage evaluations, and urban planning across large geographical areas. In regions like Korea, where real 26estate markets are rapidly evolving, valuation models based on multiple linear regression are valued for their 27simplicity and interpretability but often fall short in capturing complex market dynamics. In contrast, machine 28learning (ML) models, while addressing non-linear relationships between property characteristics and market 29values and offering superior predictive performance, are often criticized for their “black-box” nature, which 30raises concerns over interpretability in transparency-critical domains like property tax assessments and policy 31planning. To address these concerns, this study investigates the application of Explainable AI (XAI) techniques 32in the mass appraisal of residential properties in Korea, integrating XAI methods with both multiple linear 33regression and random forest models. Using SHAP (SHapley Additive exPlanations) and PFI (Permutation Fea-34ture Importance) values, the study analyzes feature importance and predictive contributions, offering insights 35into the factors driving property valuations. Additionally, a temporal analysis was conducted by segmenting 36the data into time intervals to examine how feature importance and predictive contributions evolve over time. 37By combining high predictive performance with transparent and interpretable insights, the findings dem-38onstrate that XAI can enhance the usability of both traditional and advanced automated valuation models 39(AVMs) for real-world decision-making in the Korean real estate sector.40 ■received 2 February 202541 ■accepted 28 August 202542Keywords: automatic valuation model, mass appraisal, explainable AI (XAI), SHAP (SHapley Additive exPlanations), Permutation Feature Importance (PFI), 43temporal analysis. 44 * Corresponding author. E-mail: myounggulee@konkuk.ac.kr451. Introduction 46Mass appraisal, as defined by the International Associa-47tion of Assessing Officers (IAAO), refers to “the process 48of valuing a group of properties as of a given date us-49ing common data, standardized methods, and statistical 50testing”. This method is extensively employed by local 51governments for property taxation, as it allows for the 52efficient and timely valuation of a large number of prop-53erties (IAAO, 2017). Given that property tax assessments 54must be conducted regularly—particularly in markets 55with rapid changes in economic conditions—mass ap-56praisal provides a pragmatic solution for achieving both 57cost-effectiveness and timeliness in valuation processes. 58International regulatory frameworks, such as the Basel 59II Accord established by the Basel Committee on Bank-60ing Supervision (BCBS) in 2008, further underscore the 61importance of regular property valuation. These regula-62tions require financial institutions to monitor the value 63of collateral frequently, mandating at least one valuation 64annually, with more frequent assessments recommended 65in highly volatile markets (Hong et al., 2020). The grow-66ing emphasis on accurate and timely property valuations 67reflects their critical role in maintaining financial stability. 68As a result, the demand for robust and scalable mass ap-69praisal models has significantly increased. 70Nevertheless, valuing real estate presents distinctive 71challenges due to the heterogeneous nature of properties. 72Unlike homogeneous commodities, residential properties 73are characterized by immobility, durability, and variability 74in both physical features and market conditions (McClus-75key et al., 2000). Furthermore, a multitude of factors—in-76cluding market segmentation and government interven-77tion—can influence property prices, adding complexity to 78the valuation process. The fundamental objective of valu-79ation models is to produce a reliable, credible, and cost-80effective estimate of a property’s market value at a given 81point in time. Achieving this goal is essential not only for 82property tax assessments but also for broader applications 83such as portfolio risk management, insurance valuations, 84and urban planning (Glumac & Des Rosiers, 2021).85202586Volume 2987Issue 588Pages 350–37689https://doi.org/10.3846/ijspm.2025.25137909192===== PAGE 13 (score 28, numeric-paren 0, table-words 4, var-words 14) =====939436295W.-s. Kim et al. Explainable AI-based mass appraisal: Insights from machine learning applications in Korea’s residential property market965. Results and discussions97This study employed a 5-fold cross-validation approach, 98which divides the dataset into five subsets, using four sub-99sets for training and one for validation in each iteration. 100This ensures a robust evaluation of the model’s predictive 101performance by minimizing the impact of data partitioning 102on the results. Using the approach, for the linear regres-103sion model, the results showed an R2 value of 0.892 and a 104MAPE of 0.12227. In comparison, the random forest model 105achieved an R2 value of 0.970 and a MAPE of 0.05928. 106Numerous studies have highlighted the superior predictive 107performance of random forest models compared to linear 108regression models, making this result somewhat expected 109(Hong et al., 2020). Since multicollinearity caused by corre-110lations among variables can pose a problem in both multi-111ple linear regression models and Permutation Feature Im-112portance (PFI) analysis, we examined the Variance Inflation 113Factor (VIF) scores for all variables. The analysis revealed 114that most VIF values were below 10, suggesting that 115multicollinearity does not significantly impact our results 116(see Appendix Table A2). Also, as mentioned earlier, XAI 117techniques were incorporated to enhance the interpret-118ability of the model. The mean absolute SHAP values and 119Permutation Feature Importance scores (PFI) were applied 120to analyze the importance of features in predicting prop-121erty prices and to provide insights into the relationships 122between features and the target variable. The insights de-123rived from SHAP and PFI analyses were further utilized to 124identify key drivers of property prices, offering actionable 125information to support decision-making in the real estate 126domain. The analysis was conducted using Python and rel-127evant libraries, including SHAP and scikit-learn, ensuring 128the reproducibility and transparency of the results. 1295.1. Comparison of random forest and linear 130regression methods131The Mean Absolute SHAP values and Permutation Feature 132Importance scores were calculated for both a linear regres-133sion model and a random forest model, and the results are 134presented in Tables 3 and 4. It is noteworthy that, in the lin-135ear regression model, the target variable was transformed 136by applying a logarithmic function to apartment prices in 137order to improve predictive performance. This transfor-138mation stabilizes the variance of the target variable and 139enables the model to better capture relationships between 140features and prices. However, as a consequence, the SHAP 141and PFI values derived from the linear regression model 142reflect the importance of features on the logarithmic scale 143rather than the original price scale. Therefore, the numeri-144cal values of SHAP and PFI from the linear regression mod-145el cannot be directly compared to those from the random 146forest model, which evaluates feature importance on the 147original price scale. Nonetheless, within each model, these 148metrics provide valuable insights into the relative impor-149tance of features and their contributions to the target vari-150able. Additionally, by examining the rankings and patterns 151of feature importance across both models, it is possible 152to identify which variables are consistently significant and 153how their influence may differ depending on the modeling 154approach. This analysis enables a nuanced interpretation of 155the factors driving apartment prices.156Variables157Mean158Median159Standard 160deviation161Min162Max163Number of units in the complex1641666.961658051661.0916771685040169Number of buildings in the complex17028.4217181721.441731174124175Parking lot 1761.02921771.001780.58231790.271804.53181Floor area ratio (FAR)182258.521832371840.81185721862435187Building coverage ratio (BCR)18824.91189191900.6019112192204193The tallest building height19414.08195141960.54197419846199The shortest building height20011.26201122020.50203320426205Distance to national park2061060.182071053.212080.37738820986.10792102142.47211Distance to high school212535.593213522.6092140.44332321531.88292161531.52217Distance to redevelopment area218639.331219571.5832200.659562210.002223758.56223Distance to university2243369.992253527.722260.37969122724.58662287136.5229Distance to general hospital2301037.13231965.3892320.49390323341.63272343470.83235Distance to museum236986.6872371038.932380.38086723987.49012403323.86241Distance to subway station242660.919243557.8492440.58124724547.4872462559.07247GDP (billion won)248349295.822493537432500.16251240439252446835253Economic growth rate 2543.3622553.22560.529257–1.92587.4259Land price fluctuation rate2600.007701452610.015809626239.504263–2.642752640.351297265Mortgage interest rate2666.148952675.908532680.09728372695.26332707.41544271End of Table 2272273274===== PAGE 14 (score 28, numeric-paren 1, table-words 2, var-words 16) =====275276International Journal of Strategic Property Management, 2025, 29(5), 350–376277363278Table 3. Explanatory indicators for apartment prices in the linear regression model279Linear regression model280Rank281Mean absolute SHAP282Permutation feature importance2831284Area2850.2879490 286Area2870.0270373 2882289Number of units in the complex2900.1652130 291Number of units in the complex2920.0124664 2933294Dong2950.1650627 296Dong2970.0106721 2984299Construction year3000.1363777 301Construction year3020.0071438 3035304GDP3050.0979463 306The tallest building height3070.0053472 3086309The tallest building height3100.0932182 311GDP3120.0050100 3137314Transaction year and quarter3150.0804335 316Apartment brand3170.0032981 3188319The shortest building height3200.0658740 321Transaction year and quarter3220.0030130 3239324Number of buildings3250.0492220 326The shortest building height3270.0029705 32810329Apartment brand3300.0454338 331Number of buildings3320.0020738 33311334Heating system3350.0446536 336Heating system3370.0011156 33812339Mortgage interest rate3400.0354506 341Mortgage interest rate3420.0009495 34313344Building coverage ratio (BCR)3450.0213643 346Floor area ratio (FAR)3470.0006287 34814349Floor area ratio (FAR)3500.0213598 351Building coverage ratio (BCR)3520.0004180 35315354Floor level3550.0109712 356Floor level3570.0001145 35816359Economic growth rates3600.0064508 361Parking lot3620.0000675 36317364Parking lot3650.0063428 366Economic growth rates3670.0000547 36818369Land price fluctuation rate3700.0028206 371Land price fluctuation rate3720.0000272 37319374Distance to the nearest national 375park3760.0025191 377Distance to the nearest national 378park3790.0000147 38020381Distance to the redevelopment area 0.0019281 382Distance to the nearest subway 383station3840.0000056 38521386Distance to the nearest general 387hospital3880.0013635 389Distance to the nearest national 390museum3910.0000013 39222393Distance to the nearest subway 394station3950.0011362 396Distance to the nearest university3970.0000001 39823399Distance to the nearest national 400museum4010.0009954 402Distance to the nearest high school403–0.000000640424405Distance to the nearest high school4060.0005801 407Distance to the redevelopment area –0.000003140825409Distance to the nearest university4100.0001085 411Distance to the nearest general 412hospital413(0.0000044)414Table 4. Explanatory indicators for apartment prices in the random forest model415Random forest model416Rank417Mean absolute SHAP418Permutation feature importance4191420Area42122098.10422Area4230.45935900 4242425Number of buildings42610862.90 427Number of buildings4280.25295000 4293430GDP4316397.02 432GDP4330.12923800 4344435Dong4363643.78437Dong4380.05789090 4395440The shortest building height4412576.39 442Building coverage ratio (BCR)4430.04592850 4446445Number of units in the complex4462439.17 447Number of units in the complex4480.03644470 4497450Mortgage interest rate4511378.81 452Mortgage interest rate4530.03068410 4548455Floor area ratio (FAR)4561344.20 457Floor area ratio (FAR)4580.02470100 4599460The tallest building height4611339.18 462The shortest building height4630.02383690 46410465Parking lot4661270.30 467Land price fluctuation rate4680.0216929046911470Building coverage ratio (BCR)4711152.89 472The tallest building height4730.01770560 47412475Land price fluctuation rate4761011.33 477Parking lot4780.01360550 47913480Construction year481944.63 482Economic growth rates4830.01288790 484485486===== PAGE 15 (score 18, numeric-paren 0, table-words 1, var-words 14) =====487488364489W.-s. Kim et al. Explainable AI-based mass appraisal: Insights from machine learning applications in Korea’s residential property market490Random forest model491Rank492Mean absolute SHAP493Permutation feature importance49414495Apartment brand496833.71 497Construction year4980.01247240 49915500Floor level501605.82502Floor level5030.01231590 50416505Transaction year and quarter506586.77507Transaction year and quarter5080.00925989 50917510Economic growth rates511569.85 512Apartment brand5130.00768583 51418515Distance to the redevelopment area 261.18 516Distance to the nearest national 517park5180.00480895 51919520Distance to the nearest national 521park522205.44 523Distance to the redevelopment area 0.00429915 52420525Heating system526180.99 527Distance to the nearest university5280.00326974 52921530Distance to the nearest university531113.81532Distance to the nearest high school5330.00306591 53422535Distance to the nearest subway 536station537104.85538Distance to the nearest subway 539station5400.00303488 54123542Distance to the nearest high school54392.14 544Distance to the nearest national 545museum5460.00302023 54724548Distance to the nearest national 549museum55087.55551Distance to the nearest general 552hospital5530.00265319 55425555Distance to the nearest general 556hospital55784.96558Heating system5590.00173744 560End of Table 4561The comparison between the linear regression and 562random forest models reveals both similarities and dif-563ferences in evaluating feature importance for predicting 564apartment prices. Both models identified key features, 565such as “Area”, “Number of units in the complex”, “Dong”, 566“GDP”, and “Construction year”, as highly significant, sug-567gesting that these variables have strong intrinsic relation-568ships with apartment prices regardless of the modeling 569approach. However, notable differences were observed 570in the rankings of certain variables, reflecting the distinct 571characteristics of each model.572For example, in the linear regression model, “Number 573of units in the complex” was ranked as a more important 574feature compared to “Number of buildings”. In contrast, 575the random forest model assigned higher importance 576to “Number of buildings” and “Building coverage ratio 577(BCR)”. Both variables represent aspects of the scale of 578an apartment complex, but their differing rankings reflect 579the fundamental characteristics of the two modeling ap-580proaches. “Number of units in the complex” has a more 581linear relationship with apartment prices, which aligns 582with the assumptions of the linear regression model. As 583a result, it is prioritized in the feature ranking within this 584model. On the other hand, “Number of buildings”, while 585less linearly correlated with apartment prices, captures 586more complex interactions and nonlinear relationships 587that the random forest model, based on decision-tree 588algorithms, is well-equipped to detect. Additionally, 589“Building coverage ratio (BCR)” complements “Number 590of buildings” by providing a related but distinct measure 591of the physical scale and layout of apartment complexes, 592further emphasizing the random forest model’s ability 593to account for these features. These differences suggest 594that the random forest model interprets apartment scale 595through a combination of interacting and nonlinear fac-596tors, whereas the linear regression model prioritizes sim-597pler, more direct relationships.598In the linear regression model, “Construction year” 599was ranked higher in importance compared to the ran-600dom forest model. This difference can be attributed to the 601distinct characteristics of the two modeling approaches. 602“Construction year” often has a linear or near-linear re-603lationship with apartment prices, as newer buildings are 604generally associated with higher market values due to 605better infrastructure, design, and amenities. The linear re-606gression model captures this straightforward relationship 607effectively, leading to a higher ranking for this feature. In 608contrast, the random forest model relies on decision-tree-609based partitions to capture nonlinear relationships and 610interactions among features. While “Construction year” 611may still influence apartment prices in this framework, its 612impact might be partially distributed across interactions 613with other variables, such as “Building coverage ratio 614(BCR)”, “Floor area ratio (FAR)” or “Number of units in the 615complex”. As a result, its individual importance is diluted, 616leading to a lower ranking in the random forest model. 617Moreover, the random forest model may prioritize features 618that have clear thresholds or significant nonlinearity, which 619“Construction year” does not exhibit as strongly. 620Interestingly, in the linear regression model, “Tallest 621building height” was ranked higher in importance com-622pared to “Shortest building height”. Conversely, in the ran-623dom forest model, the ranking was reversed, with “Short-624est building height” being more important. The linear 625regression model prioritizes “Tallest building height” due 626to its more direct and linear relationship with apartment 627prices, as taller buildings often signify premium develop-628ments. On the other hand, the random forest model might 629