Compressive strength prediction of slag-blended concrete: A novel interpretable CGLRXS-WV-SHAP framework

The mainstream approach for predicting the compressive strength of slag concrete currently relies on "black-box" machine learning models. Such models are prone to overfitting when handling complex nonlinear relationships and lack interpretability. To address this issue, the present study innovatively proposes an interpretable prediction framework integrating multiple techniques, termed CGLRXS-WV-SHAP, which aims to overcome the limitations of overfitting susceptibility and lack of interpretability inherent in black-box models . This framework employs WV strategy to ensemble six base models including CatBoost(CB), Gradient Boosting Decision Tree(GBDT), LightGBM(LGBM), RandomForest(RF), XGBoost(XGB) and Symbolic Regression(SR). Based on 1030 slag concrete sample data, the Tree-structured Parzen Estimator-based Bayesian Optimization (TPE BO) method is employed to perform parameter tuning for each base model, and the optimal weight allocation scheme for different models in model fusion is determined through random search, while the SHAP tool is introduced to achieve global and local interpretation of model prediction mechanisms. Experimental results indicate that after weighted fusion, test set R ² reaches 0.93, with the difference between training set and test set R ² only 0.02, significantly reduced compared to the initial difference of 0.09 for the single XGB model, with overfitting risk substantially decreased; SHAP global feature importance analysis reveals that cement dosage (28.4%), curing age (25.6%), and slag dosage (16.1%) are the three key factors influencing compressive strength. In practical engineering applications, this framework provides an interpretable method with strong generalization capability and robustness for strength prediction of concrete containing industrial waste materials, and can support engineering design and performance optimization of low-pollution concrete structures.