An explainable model for the mass appraisal of residences: The application of tree-based Machine Learning algorithms and interpretation of value determinants

Abstract

In the mass appraisal of properties, Machine Learning (ML) algorithms have produced effective and promising results. Analysts use various algorithms to train their models with limited data and make appraisals on large data sets. However, research into which value determinants the models take into account when appraising values is insufficient. This research looks at how eXplainable Artificial Intelligence (XAI) methods can be integrated with mass real estate appraisal studies. Experimental studies were carried out on a data set containing 1002 samples and 43 independent variables. Tree-based ML regressors, namely Random Forest, XGBoost, LightGBM, and Gradient Boosting, were used for training the predictive models. The performance of these regressors was compared with that of classical multiple regression analysis. The Permutation Feature Importance (PFI) technique was used for the selection of the variables that contributed the most to the training of the models. Models retrained with selected variables were locally interpreted using the SHapley Additive eXplanations (SHAP) method. In this way, it was possible to observe the value determinants that contribute to the price estimation of each real estate sample. This study demonstrates that XAI approaches should be integrated into mass real estate valuation systems specifically, and into urban and housing research more generally, helping analysts and scholars to explain their models more transparently. The outcomes of this study can be a harbinger for analysts and scholars who wish to explain their models more transparently. Last but not least, this study advocates the use of tree-based ML algorithms since they not only allow us to implement XAI approaches but also outperform the stand-alone ML regressors.