Machine learning regression and classification algorithms utilised for strength prediction of OPC/by-product materials improved soils

In this study, stand-alone machine (ML) models (Bayesian regressor (BLR), least square linear regressor (REG), artificial neural networks (ANN), and logistic regression (LR)), tree-ensemble ML models (boosted decision tree (BDT), random decision forest (RDF) decision jungle (DJ)) and meta-ensemble ML models (voting (VE) and stacking (SE)) are applied to predict the strength of different soils improved by part substitution of OPC with PFA and GGBS in various combinations and proportions. Multiclass elements of these proposed ML models are also deployed to provide analysis across multiple cross-validation methods. Results of regression analysis indicated higher statistical variance of OPC-substituted predictor variables compared to soils improved by OPC alone when using both stand-alone and tree-based algorithms. On average, the REG model produced strength predictions with higher accuracy (RMSE of 0.39 and R-2 of 0.86) compared to ANN (RMSE of 0.44 and R-2 of 0.82), but with comparatively lower accuracy compared to tree-based models (average RMSE of 0.33 and R-2 of 0.90) and meta-ensemble models (average RMSE of 0.06 and R-2 of 0.91). For ML classification, multiclass neural network algorithm (mANN) produced higher accuracy (0.78), precision (0.67) and rate of recall (0.67) compared to tree based models but fell short to meta-ensemble models (average accuracy of 0.80, precision of 0.70 and recall of 0.71). Diagnostic tests across different validation methods indicated better performance of the VE model compared to its SE ML counterpart when adopting the train-validation split technique. Overall, the ensemble methods were more versatile on regression and multiclass classification problems because they aggregated multiple learners to provide robust predictions. (C) 2021 Elsevier Ltd. All rights reserved.