Meta-Probabilistic Stacking for Eco-Driving Behavior Profiling

Abstract

Automated driver behavior profiling is a strategically important problem within intelligent transportation systems, road safety engineering, and fleet-based emissions governance. This study presents a Heterogeneous Out-of-Fold (OOF) Stacking Ensemble augmented with inter-model probabilistic disagreement as a first-class meta-feature. Five architecturally diverse base learners — Gradient Boosting, LightGBM, Random Forest, AdaBoost, and ResNet-Dense — are each optimized via Bayesian hyperparameter search (Optuna) before integration within a 5-fold stratified OOF stacking protocol. At the meta-classification layer, an XGBoost meta-learner ingests both the 15-dimensional base-learner probability matrix and three probability-dispersion features computed as the per-class standard deviation of base-learner outputs, enabling uncertainty-conditioned arbitration at classification boundaries. Experiments on a publicly available eco-driving benchmark dataset comprising 30,000 generated trip-level records with variables representative of telematics/OBD-related driving metrics demonstrate 87.77 % accuracy, weighted F1-score of 0.8771, and macro AUC-ROC of 0.9722. These results surpass the strongest individual baseline (Gradient Boosting, 87.32 % post-tuning) and both preceding stacking generations (Generation 1: 87.28 %; Generation 2: 87.48 %). SHapley Additive exPlanations (SHAP) analysis confirms that the engineered probability-dispersion features have relatively modest global importance (ranked 13th–18th) but provide directionally informative contributions for high-disagreement, boundary-adjacent samples, validating the theoretical motivation for uncertainty-conditioned arbitration. McNemar’s test yields p = 0.023 (χ2(1) = 5.14), confirming statistical significance; the 95 % Wilson interval is [86.91%, 88.57%] (N = 6,000). The Expected Calibration Error (ECE) of the proposed ensemble (0.0107) is 45.8 % lower than the best individual model (0.0198), demonstrating superior probability calibration alongside improved discriminative accuracy. These findings support applicability to fleet telematics, pay-how-you-drive insurance scoring, and near-real-time periodic coaching platforms, pending cross-population validation.