Machine Learning interpretável na avaliação de imóveis em massa: aplicação do SHAP

Abstract

This study aims to advance the use of machine learning algorithms in mass real estate appraisal context by exploring and validating tools designed to enhance model interpretability and explainability. Using a dataset of 6,660 apartments from Fortaleza, Brazil, provided by the Municipal Finance Department, the performance of a traditional multiple linear regression model was compared with that of modern machine learning algorithms, namely XGBoost and Random Forest. XGBoost achieved the best performance across all evaluated metrics (MAPE, RMSE, R², COD, and PRD). To address the challenge of interpretability, SHAP (Shapley Additive Explanations) method was employed, enabling the assessment of each variable’s individual contribution and a global understanding of relationships captured by the model. Comparison between partial dependence plots from the classical model and SHAP-based plots for XGBoost revealed a non-linear relationship between building age and unit price, as well as an interaction between age and neighborhood income levels. Diverse effects were also observed in binary variables. The local analysis showed that the influence of predictors varies substantially across properties with different characteristics. SHAP proved valuable for enhancing transparency and improving model interpretability for both technical and public administration purposes, contributing to greater tax fairness and better communication with taxpayers. Waterfall plots, in particular, demonstrated effectiveness in explaining individual predictions without requiring in-depth technical knowledge of the algorithms. This study highlights the feasibility of integrating explainable artificial intelligence into fiscal mass appraisal workflows and recommends its adoption to strengthen and expand the use of tree-based algorithms in property taxation.