From Global Accuracy to Reliable Decisions: A Reliability-Aware Selective Prediction Framework for Maximum Dry Density of Expansive Soils

Reliable estimation of maximum dry density (MDD) is essential for compaction quality control of expansive soils. However, existing machine learning approaches mainly focus on overall accuracy metrics and rarely assess the reliability of individual predictions. This study develops a reliability-aware selective prediction framework integrating regression prediction with sample-level reliability assessment. The framework was developed using 311 expansive soil samples with five soil properties, including Ret#4, Pass#200, LL, PI, and Gs. Twelve regression models were compared through five-fold cross-validation, and Friedman–Nemenyi analysis was applied to select the optimal model. Extra Trees (ET) was selected for MDD prediction, while a Random Forest (RF) classifier was developed to assess prediction reliability using soil properties, predicted MDD values, and local feature-space information. Reliability labels were generated from cross-validation errors and local stability constraints. On the independent test set, the selective prediction strategy reduced the RMSE of retained predictions from 72.1 kg/m3 to 20.0 kg/m3 compared with full-coverage prediction. SHAP analysis further revealed the contribution of soil properties and prediction characteristics to MDD estimation and reliability assessment. The results show that the proposed framework can identify more reliable MDD predictions and avoid treating all model outputs as equally trustworthy.