Interpretable machine learning model for analyzing the effect of building characteristics on cost with Extreme Gradient Boosting and Shapley Additive Explanations

Abstract

A building consists of many components, each with its own financial impact. A clear evaluation of these impacts is possible through the analysis of the components that make up construction costs. This study aims to evaluate and interpret the combined use of construction cost components for different scenarios using machine learning models. For this purpose, the Extreme Gradient Boosting, Light Gradient Boosting Machine, and Shapley Additive Explanations algorithms were applied. Additionally, the Extreme Gradient Boosting algorithm was enhanced using the Grid Search and Random Search optimization algorithms. Specifically, the model developed with Extreme Gradient Boosting using the Grid Search method increased model performance. Performance metrics such as R-Squared, Nash-Sutcliffe Efficiency, Mean Absolute Error, Root Mean Square Error, and Mean Absolute Percentage Error were used for the numerical evaluation of the results. With the help of optimization techniques, the artificial intelligence model was improved by rates ranging from 1.5 percent to 5 percent based on these performance metrics. As a result, the study demonstrates that different representations of the same construction data can influence both prediction performance and model interpretability. A novel scenario-based framework was developed by grouping input variables into Height, Architectural, and Structural feature categories and reintroducing their aggregated representations into the prediction models. The proposed framework contributes to the planning of accurate cost estimates and more efficient resource utilization by providing insights into how alternative feature-grouping strategies affect both construction cost prediction performance and explainable machine learning interpretations.