Evaluating the Sensitivity of Machine Learning Models to Data Preprocessing Technique in Concrete Compressive Strength Estimation.

This study rigorously examines the impact of various data preprocessing techniques on the accuracy of machine learning models in predicting concrete's compressive strength. It develops ten regression models under nine distinct preprocessing scenarios, including normalization, standardization, principal component analysis (PCA), and polynomial features, utilizing a comprehensive dataset featuring normal and high-strength performances. The results reveal that using polynomial features and kernel PCA significantly enhanced model performance, with R values soaring to 93.27 and 94.65% during training and 88.51 and 88.77% during testing, respectively. This indicates their strong ability to capture the hidden nonlinear relationships within data. Conversely, discretization exhibited the least effectiveness, with the highest normalized root mean square error values of 14.2 (training) and 16.8 (testing) and normalized mean absolute error values of 11.6 (training) and 13.6 (testing), suggesting a potential loss of essential data granularity. Additionally, the study found that machine learning techniques generally surpassed traditional regression models, with higher R values being a consistent trend. These findings offer a nuanced understanding of the importance of preprocessing choice in concrete strength prediction and provide valuable insights for the concrete industry and data scientists, emphasizing the critical role of data preprocessing in achieving optimal model accuracy in materials science. [ABSTRACT FROM AUTHOR]