Towards Green AI: The Auto-Datavise Framework for Energy-Aware and Interpretable Machine Learning

Automated machine learning (AutoML) has broadened access to ML, yet significant gaps remain in achieving fully autonomous, efficient, and interpretable systems. Auto-Datavise is presented as a glass-box AutoML framework for tabular data which unifies automated feature synthesis, wrapper-based feature selection, and Bayesian hyperparameter optimization within reproducible scikit-learn pipelines. Interpretability is provided via SHapley Additive exPlanations (SHAP), permutation importance, Partial- Dependence Plots (PDP), learning curves, and residual diagnostics, whereas energy/CO 2 emissions are tracked to enable Green-AI reporting. Multi-objective tuning jointly maximizes predictive performance while minimizing runtime, with Pareto analysis employed to characterize performance-cost trade-offs. On the Building Energy Efficiency dataset (Case-1), CatBoost attained near-ceiling accuracy (R 2 = 0.9989) with favorable performance-cost trade-offs. On Concrete Strength (Case-2), SHAP ranked Age and Cement as dominant drivers, aligning with engineering knowledge. On a noisy Engineering-Salary dataset (Case- 3), the raw baseline (R 2 = 0.1962) improved to R 2 = 0.260 after automated cleaning and RFECV, with tighter predicted-vs-observed alignment and diminished overfitting. Robustness against H2O AutoML and TPOT under Gaussian noise (σ = 0.05·σ), adversarial-type flips (k = 3 ε = 0.1), and mean shift (+0.5·σ) revealed small ΔR 2 . On Case-3 (noisy data), Auto-Datavise consistently exceeded TPOT and remained close to H2O. TPOT collapsed under mean shift, while Auto-Datavise degraded least. On Case-1, Auto- Datavise outperformed the comparison frameworks under Gaussian noise and adversarial-type feature perturbations. Under mean shift it trailed H2O by only 2.5% (ordering: H2O > Auto-Datavise > TPOT).