Performance analysis of Gradient Boosting Regressor using K-fold and group K-fold for predicting permanent deformation in tropical soils

This study conducts a comparative analysis of k-fold and group k-fold cross-validation techniques for predicting permanent deformation in tropical soils using the Gradient Boosting Regressor (GBR). Using a dataset of 26 Brazilian soil samples, subjected to repeated load triaxial tests, gradation, compaction, and Miniature, Compacted Tropical (MCT) classification analyses, we developed two predictive models: GBR with k-fold and GBR with group k-fold. Data were divided 5into training (80%) and testing (20%) sets, with a 10-fold scheme for hyperparameter optimization, 6training, and validation. Results showed that k-fold induced overfitting, negatively affecting test set generalization (R2= 0.7163, RMSEtest = 0.4070). In contrast, group k-fold demonstrated robust performance (R2= 0.9060, RMSEtest = 0.2342). Feature importance analysis revealed that test considering groups of correlated observations alters variable importance. Variables σd, γdmax , d′, and e′ were consistently important, while OMC and σ3 gained relevance in the group k-fold model. Thus, group k-fold is recommended for scenarios with correlated observations to enhance model generalization. The GBR model with group k-fold successfully predicted the permanent deformation of tropical soils.