Impact of feature engineering on machine learning-based prediction of concrete compressive strength
Asian Journal of Civil Engineering, cilt.27, sa.11, ss.5159-5179, 2026 (Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 27 Sayı: 11
- Basım Tarihi: 2026
- Doi Numarası: 10.1007/s42107-026-01763-y
- Dergi Adı: Asian Journal of Civil Engineering
- Derginin Tarandığı İndeksler: Scopus, zbMATH, Materials Science & Engineering Collection (ProQuest), Technology Collection (ProQuest)
- Sayfa Sayıları: ss.5159-5179
- Anahtar Kelimeler: Concrete compressive strength, Domain knowledge, Feature engineering, Machine learning, SHAP analysis
- Atatürk Üniversitesi Adresli: Evet
Özet
Accurate prediction of concrete compressive strength is essential for mix design optimization and structural reliability assessment. However, the extent to which explicit feature engineering improves prediction performance beyond the choice of machine learning algorithm remains unclear. This study systematically investigates this question using the widely used Concrete Compressive Strength dataset and three data representation scenarios: (i) a baseline scenario containing raw input variables (raw scenario), (ii) an intermediate scenario involving ratio-based transformations (basic ratios scenario), and (iii) an advanced scenario incorporating engineered features derived from established physical relationships in concrete technology, such as the water/binder ratio and the logarithmically transformed curing age (engineered ratios scenario). Prediction performance was compared across Random Forest, XGBoost, LightGBM, and Support Vector Regression (SVR) models using a 2020 nested cross-validation scheme with grouped folds to prevent information leakage between specimens sharing the same mix design, and the statistical significance of performance differences was assessed using the Bonferroni-corrected Wilcoxon signed-rank test. The best-performing configuration (LightGBM under the engineered ratios scenario) achieved and Root Mean Square Error (RMSE) = 4.84 MPa. Ratio-based transformations alone did not yield a statistically significant improvement over raw inputs for any model. The engineered ratios scenario produced a statistically significant improvement only for Random Forest (increased from 0.869 to 0.893,); the corresponding improvements for XGBoost, LightGBM, and SVR were not statistically significant, despite modest increases in mean. SHapley Additive exPlanations (SHAP)-based interpretability analyses showed that the engineered water/binder ratio and logarithmically transformed curing age were consistently among the most influential features across all models, with dependence patterns consistent with established relationships in concrete technology, such as Abrams’ Law. These findings indicate that the benefit of physically motivated feature engineering is model-dependent rather than universal, and underscore the importance of leakage-aware cross-validation design when evaluating feature engineering strategies for concrete strength prediction.