Comparative study of energy and accuracy in spatio-temporal models for engagement classification

Abstract

Energy consumption is a growing concern in deep learning, motivating the Green AI paradigm, where models are evaluated not only based on predictive performance but also on energy efficiency. Most existing evaluations focus on spatial tasks and do not fully capture the computational and temporal costs of video-based workflows. This study presents a systematic energy evaluation of spatio-temporal deep learning for video-based student engagement classification. Using the DAiSEE dataset, we compare EfficientNet variants (B0–B7) as frame-level feature extractors and model temporal dynamics with LSTMs on fixed-length sequences. GPU power consumption is measured during training using nvidia-smi, and energy efficiency is quantified with the Kappa Energy Index (KEI), defined as Cohen's Kappa divided by energy consumption (kWh). The results show a clear trade-off between accuracy and energy: EfficientNet-B7 achieves the highest accuracy (0.62) but incurs the highest energy cost (≈5 kWh), resulting in a low KEI, while EfficientNet-B0 achieves competitive accuracy (0.59) with the highest KEI (2.147) due to its low energy consumption (≈0.38 kWh). EfficientNet-B3 strikes a favorable balance (accuracy ≈ 0.61, KEI = 0.747), outperforming larger models under resource constraints. These findings suggest that deeper models do not always maximize energy efficiency, and the KEI value provides a practical metric to guide the selection of energy-efficient models.