Enhancing the accuracy and interpretability of real estate price predictions using machine learning methods

Abstract

Relevance: stems from the increasing need for machine learning models in real estate that are not only accurate but also interpretable, since practical applications require a clear understanding of the factors influencing predictions, particularly in decision-making related to urban development and property valuation. Aim: the aim of the article is to develop and evaluate an approach that simultaneously improves predictive performance and enhances the interpretability of machine learning models when working with heterogeneous data sources. Objectives: the objectives include analysing multiple feature groups, namely textual property descriptions, spatial indicators such as distances to key infrastructure objects, and visual characteristics derived from satellite-based night-time illumination data, as well as assessing their combined impact on model performance. Methods: the study is based on gradient boosting over decision trees using the Light Gradient Boosting Machine algorithm, the construction of spatial features through distance-based metrics, text processing techniques, and interpretation tools based on Shapley values and partial dependence analysis to reveal feature influence. Scientific novelty: the novelty lies in integrating heterogeneous features of different origin within a unified modelling framework and combining complementary interpretation techniques to identify nonlinear relationships and interaction effects. Practical significance: the proposed approach can be applied in automated valuation systems, urban analytics, and decision-support tools, providing a more transparent understanding of price formation mechanisms. Results: the results show that the proposed modelling approach based on the Light Gradient Boosting Machine provides high predictive performance. Starting from the baseline feature set, the model achieves an RMSE of 374 and an R² of 0.818. The integration of heterogeneous feature groups further improves the model performance, reducing the RMSE to 348 and increasing the R² to 0.839. It is also evident that textual and visual features play a noticeable role, as they help capture nonlinear patterns and threshold-like effects that are difficult to detect otherwise. Conclusions: overall, the proposed approach not only increases predictive accuracy but also makes the model behavior easier to interpret, leading to a more transparent and reliable analysis of real estate prices and supporting its use in practical applications.