# RECORD ee4a73efc5ad67b4400db23b09eeacbeb4812f86 # TITLE: Explainable AI-based mass appraisal: Insights from machine learning applications in Korea’s residential property market # YEAR: 2025.0 VENUE: International Journal of Strategic Property Management DOI: 10.3846/ijspm.2025.25137 # AUTHORS: Woosung Kim; Insu Choi; Myounggu Lee # PAGES: 27 CANDIDATE TABLE PAGES (1-based): [13, 14] # 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. ===== PAGE 1 (score 9, numeric-paren 0, table-words 0, var-words 9) ===== This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Copyright © 2025 The Author(s). Published by Vilnius Gediminas Technical University ISSN: 1648-715X / eISSN: 1648-9179 INTERNATIONAL JOURNAL OF STRATEGIC PROPERTY MANAGEMENT EXPLAINABLE AI-BASED MASS APPRAISAL: INSIGHTS FROM MACHINE LEARNING APPLICATIONS IN KOREA’S RESIDENTIAL PROPERTY MARKET Woo-sung KIM   1, Insu CHOI   2, Myounggu LEE   1* 1 Department of Business Administration, Konkuk University, Seoul, South Korea 2 Division of Business Administration, Dongduk Women’s University, Seoul, South Korea Article History: 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 Fea- ture 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 dem- onstrate 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. ■received 2 February 2025 ■accepted 28 August 2025 Keywords: automatic valuation model, mass appraisal, explainable AI (XAI), SHAP (SHapley Additive exPlanations), Permutation Feature Importance (PFI), temporal analysis. * Corresponding author. E-mail: myounggulee@konkuk.ac.kr 1. Introduction Mass appraisal, as defined by the International Associa- tion of Assessing Officers (IAAO), refers to “the process of valuing a group of properties as of a given date us- ing common data, standardized methods, and statistical testing”. This method is extensively employed by local governments for property taxation, as it allows for the efficient and timely valuation of a large number of prop- erties (IAAO, 2017). Given that property tax assessments must be conducted regularly—particularly in markets with rapid changes in economic conditions—mass ap- praisal provides a pragmatic solution for achieving both cost-effectiveness and timeliness in valuation processes. International regulatory frameworks, such as the Basel II Accord established by the Basel Committee on Bank- ing Supervision (BCBS) in 2008, further underscore the importance of regular property valuation. These regula- tions require financial institutions to monitor the value of collateral frequently, mandating at least one valuation annually, with more frequent assessments recommended in highly volatile markets (Hong et al., 2020). The grow- ing emphasis on accurate and timely property valuations reflects their critical role in maintaining financial stability. As a result, the demand for robust and scalable mass ap- praisal models has significantly increased. Nevertheless, valuing real estate presents distinctive challenges due to the heterogeneous nature of properties. Unlike homogeneous commodities, residential properties are characterized by immobility, durability, and variability in both physical features and market conditions (McClus- key et al., 2000). Furthermore, a multitude of factors—in- cluding market segmentation and government interven- tion—can influence property prices, adding complexity to the valuation process. The fundamental objective of valu- ation models is to produce a reliable, credible, and cost- effective estimate of a property’s market value at a given point in time. Achieving this goal is essential not only for property tax assessments but also for broader applications such as portfolio risk management, insurance valuations, and urban planning (Glumac & Des Rosiers, 2021). 2025 Volume 29 Issue 5 Pages 350–376 https://doi.org/10.3846/ijspm.2025.25137 ===== PAGE 13 (score 28, numeric-paren 0, table-words 4, var-words 14) ===== 362 W.-s. Kim et al. Explainable AI-based mass appraisal: Insights from machine learning applications in Korea’s residential property market 5. Results and discussions This study employed a 5-fold cross-validation approach, which divides the dataset into five subsets, using four sub- sets for training and one for validation in each iteration. This ensures a robust evaluation of the model’s predictive performance by minimizing the impact of data partitioning on the results. Using the approach, for the linear regres- sion model, the results showed an R2 value of 0.892 and a MAPE of 0.12227. In comparison, the random forest model achieved an R2 value of 0.970 and a MAPE of 0.05928. Numerous studies have highlighted the superior predictive performance of random forest models compared to linear regression models, making this result somewhat expected (Hong et al., 2020). Since multicollinearity caused by corre- lations among variables can pose a problem in both multi- ple linear regression models and Permutation Feature Im- portance (PFI) analysis, we examined the Variance Inflation Factor (VIF) scores for all variables. The analysis revealed that most VIF values were below 10, suggesting that multicollinearity does not significantly impact our results (see Appendix Table A2). Also, as mentioned earlier, XAI techniques were incorporated to enhance the interpret- ability of the model. The mean absolute SHAP values and Permutation Feature Importance scores (PFI) were applied to analyze the importance of features in predicting prop- erty prices and to provide insights into the relationships between features and the target variable. The insights de- rived from SHAP and PFI analyses were further utilized to identify key drivers of property prices, offering actionable information to support decision-making in the real estate domain. The analysis was conducted using Python and rel- evant libraries, including SHAP and scikit-learn, ensuring the reproducibility and transparency of the results. 5.1. Comparison of random forest and linear regression methods The Mean Absolute SHAP values and Permutation Feature Importance scores were calculated for both a linear regres- sion model and a random forest model, and the results are presented in Tables 3 and 4. It is noteworthy that, in the lin- ear regression model, the target variable was transformed by applying a logarithmic function to apartment prices in order to improve predictive performance. This transfor- mation stabilizes the variance of the target variable and enables the model to better capture relationships between features and prices. However, as a consequence, the SHAP and PFI values derived from the linear regression model reflect the importance of features on the logarithmic scale rather than the original price scale. Therefore, the numeri- cal values of SHAP and PFI from the linear regression mod- el cannot be directly compared to those from the random forest model, which evaluates feature importance on the original price scale. Nonetheless, within each model, these metrics provide valuable insights into the relative impor- tance of features and their contributions to the target vari- able. Additionally, by examining the rankings and patterns of feature importance across both models, it is possible to identify which variables are consistently significant and how their influence may differ depending on the modeling approach. This analysis enables a nuanced interpretation of the factors driving apartment prices. Variables Mean Median Standard deviation Min Max Number of units in the complex 1666.96 805 1.09 7 5040 Number of buildings in the complex 28.42 8 1.44 1 124 Parking lot 1.0292 1.00 0.5823 0.27 4.53 Floor area ratio (FAR) 258.52 237 0.81 72 2435 Building coverage ratio (BCR) 24.91 19 0.60 12 204 The tallest building height 14.08 14 0.54 4 46 The shortest building height 11.26 12 0.50 3 26 Distance to national park 1060.18 1053.21 0.377388 86.1079 2142.47 Distance to high school 535.593 522.609 0.443323 31.8829 1531.52 Distance to redevelopment area 639.331 571.583 0.65956 0.00 3758.56 Distance to university 3369.99 3527.72 0.379691 24.5866 7136.5 Distance to general hospital 1037.13 965.389 0.493903 41.6327 3470.83 Distance to museum 986.687 1038.93 0.380867 87.4901 3323.86 Distance to subway station 660.919 557.849 0.581247 47.487 2559.07 GDP (billion won) 349295.82 353743 0.16 240439 446835 Economic growth rate 3.362 3.2 0.529 –1.9 7.4 Land price fluctuation rate 0.00770145 0.0158096 39.504 –2.64275 0.351297 Mortgage interest rate 6.14895 5.90853 0.0972837 5.2633 7.41544 End of Table 2 ===== PAGE 14 (score 28, numeric-paren 1, table-words 2, var-words 16) ===== International Journal of Strategic Property Management, 2025, 29(5), 350–376 363 Table 3. Explanatory indicators for apartment prices in the linear regression model Linear regression model Rank Mean absolute SHAP Permutation feature importance 1 Area 0.2879490 Area 0.0270373 2 Number of units in the complex 0.1652130 Number of units in the complex 0.0124664 3 Dong 0.1650627 Dong 0.0106721 4 Construction year 0.1363777 Construction year 0.0071438 5 GDP 0.0979463 The tallest building height 0.0053472 6 The tallest building height 0.0932182 GDP 0.0050100 7 Transaction year and quarter 0.0804335 Apartment brand 0.0032981 8 The shortest building height 0.0658740 Transaction year and quarter 0.0030130 9 Number of buildings 0.0492220 The shortest building height 0.0029705 10 Apartment brand 0.0454338 Number of buildings 0.0020738 11 Heating system 0.0446536 Heating system 0.0011156 12 Mortgage interest rate 0.0354506 Mortgage interest rate 0.0009495 13 Building coverage ratio (BCR) 0.0213643 Floor area ratio (FAR) 0.0006287 14 Floor area ratio (FAR) 0.0213598 Building coverage ratio (BCR) 0.0004180 15 Floor level 0.0109712 Floor level 0.0001145 16 Economic growth rates 0.0064508 Parking lot 0.0000675 17 Parking lot 0.0063428 Economic growth rates 0.0000547 18 Land price fluctuation rate 0.0028206 Land price fluctuation rate 0.0000272 19 Distance to the nearest national park 0.0025191 Distance to the nearest national park 0.0000147 20 Distance to the redevelopment area 0.0019281 Distance to the nearest subway station 0.0000056 21 Distance to the nearest general hospital 0.0013635 Distance to the nearest national museum 0.0000013 22 Distance to the nearest subway station 0.0011362 Distance to the nearest university 0.0000001 23 Distance to the nearest national museum 0.0009954 Distance to the nearest high school –0.0000006 24 Distance to the nearest high school 0.0005801 Distance to the redevelopment area –0.0000031 25 Distance to the nearest university 0.0001085 Distance to the nearest general hospital (0.0000044) Table 4. Explanatory indicators for apartment prices in the random forest model Random forest model Rank Mean absolute SHAP Permutation feature importance 1 Area 22098.10 Area 0.45935900 2 Number of buildings 10862.90 Number of buildings 0.25295000 3 GDP 6397.02 GDP 0.12923800 4 Dong 3643.78 Dong 0.05789090 5 The shortest building height 2576.39 Building coverage ratio (BCR) 0.04592850 6 Number of units in the complex 2439.17 Number of units in the complex 0.03644470 7 Mortgage interest rate 1378.81 Mortgage interest rate 0.03068410 8 Floor area ratio (FAR) 1344.20 Floor area ratio (FAR) 0.02470100 9 The tallest building height 1339.18 The shortest building height 0.02383690 10 Parking lot 1270.30 Land price fluctuation rate 0.02169290 11 Building coverage ratio (BCR) 1152.89 The tallest building height 0.01770560 12 Land price fluctuation rate 1011.33 Parking lot 0.01360550 13 Construction year 944.63 Economic growth rates 0.01288790 ===== PAGE 15 (score 18, numeric-paren 0, table-words 1, var-words 14) ===== 364 W.-s. Kim et al. Explainable AI-based mass appraisal: Insights from machine learning applications in Korea’s residential property market Random forest model Rank Mean absolute SHAP Permutation feature importance 14 Apartment brand 833.71 Construction year 0.01247240 15 Floor level 605.82 Floor level 0.01231590 16 Transaction year and quarter 586.77 Transaction year and quarter 0.00925989 17 Economic growth rates 569.85 Apartment brand 0.00768583 18 Distance to the redevelopment area 261.18 Distance to the nearest national park 0.00480895 19 Distance to the nearest national park 205.44 Distance to the redevelopment area 0.00429915 20 Heating system 180.99 Distance to the nearest university 0.00326974 21 Distance to the nearest university 113.81 Distance to the nearest high school 0.00306591 22 Distance to the nearest subway station 104.85 Distance to the nearest subway station 0.00303488 23 Distance to the nearest high school 92.14 Distance to the nearest national museum 0.00302023 24 Distance to the nearest national museum 87.55 Distance to the nearest general hospital 0.00265319 25 Distance to the nearest general hospital 84.96 Heating system 0.00173744 End of Table 4 The comparison between the linear regression and random forest models reveals both similarities and dif- ferences in evaluating feature importance for predicting apartment prices. Both models identified key features, such as “Area”, “Number of units in the complex”, “Dong”, “GDP”, and “Construction year”, as highly significant, sug- gesting that these variables have strong intrinsic relation- ships with apartment prices regardless of the modeling approach. However, notable differences were observed in the rankings of certain variables, reflecting the distinct characteristics of each model. For example, in the linear regression model, “Number of units in the complex” was ranked as a more important feature compared to “Number of buildings”. In contrast, the random forest model assigned higher importance to “Number of buildings” and “Building coverage ratio (BCR)”. Both variables represent aspects of the scale of an apartment complex, but their differing rankings reflect the fundamental characteristics of the two modeling ap- proaches. “Number of units in the complex” has a more linear relationship with apartment prices, which aligns with the assumptions of the linear regression model. As a result, it is prioritized in the feature ranking within this model. On the other hand, “Number of buildings”, while less linearly correlated with apartment prices, captures more complex interactions and nonlinear relationships that the random forest model, based on decision-tree algorithms, is well-equipped to detect. Additionally, “Building coverage ratio (BCR)” complements “Number of buildings” by providing a related but distinct measure of the physical scale and layout of apartment complexes, further emphasizing the random forest model’s ability to account for these features. These differences suggest that the random forest model interprets apartment scale through a combination of interacting and nonlinear fac- tors, whereas the linear regression model prioritizes sim- pler, more direct relationships. In the linear regression model, “Construction year” was ranked higher in importance compared to the ran- dom forest model. This difference can be attributed to the distinct characteristics of the two modeling approaches. “Construction year” often has a linear or near-linear re- lationship with apartment prices, as newer buildings are generally associated with higher market values due to better infrastructure, design, and amenities. The linear re- gression model captures this straightforward relationship effectively, leading to a higher ranking for this feature. In contrast, the random forest model relies on decision-tree- based partitions to capture nonlinear relationships and interactions among features. While “Construction year” may still influence apartment prices in this framework, its impact might be partially distributed across interactions with other variables, such as “Building coverage ratio (BCR)”, “Floor area ratio (FAR)” or “Number of units in the complex”. As a result, its individual importance is diluted, leading to a lower ranking in the random forest model. Moreover, the random forest model may prioritize features that have clear thresholds or significant nonlinearity, which “Construction year” does not exhibit as strongly. Interestingly, in the linear regression model, “Tallest building height” was ranked higher in importance com- pared to “Shortest building height”. Conversely, in the ran- dom forest model, the ranking was reversed, with “Short- est building height” being more important. The linear regression model prioritizes “Tallest building height” due to its more direct and linear relationship with apartment prices, as taller buildings often signify premium develop- ments. On the other hand, the random forest model might