Benchmarking Energy Efficiency of Supervised Machine Learning Models on Multi-Domain Classification Datasets

Machine learning should be judged by how well it predicts, and computational resources are not accounted for in predictive accuracy. Given the growing emphasis on energy consumption and resource efficiency, decision-supporting frameworks should go beyond accuracy. This study presents an energy-based benchmarking approach for supervised learning models. Ten classical algorithms were evaluated on three textual and tabular datasets. The energy consumption of preprocessing, training, and inference was monitored with Intel RAPL via pyRAPL along with the runtime, peak memory usage, and predictive performance statistics (accuracy, precision, recall, F1-score, and AUC). Experiments were conducted in a controlled CPU-based environment to ensure comparability. The computational role of this feature is found to be appreciably diverse. Results show that Random Forest achieved the highest overall balance between predictive performance and efficiency (CI = 0.950, PPI = 0.907), while Logistic Regression provided a competitive trade-off (CI = 0.905, EI = 0.998). Gaussian Naïve Bayes was the most energy-efficient model with a mean energy consumption of 127 J, whereas Support Vector Classifier (SVC) incurred the highest computational cost, consuming 45,758 J and requiring 3925 s on average. The Pareto analysis identified Random Forest, Logistic Regression, Passive Aggressive, and Decision Tree as non-dominated solutions. These findings demonstrate that accuracy alone can be misleading for model evaluation and that integrating energy, runtime, and memory metrics enables more sustainable and resource-aware machine learning model selection. The proposed framework provides practical guidance for Green AI, Tiny Machine Learning (TinyML), edge computing, and other resource-constrained deployment environments.