A Convolutional Neural Network Bagging Ensemble for Human Emotion Classification using the AffectNet Dataset

Facial emotion recognition remains a challenging task due to the complexity of human expression and class imbalance in real-world datasets. This paper proposes a Bagging Ensemble framework for emotion classification using the manually annotated subset of the AffectNet dataset. The ensemble integrates three diverse Convolutional Neural Network (CNN) architectures, ResNet50, EfficientNet-B0, and MobileNetV2, each trained on a unique bootstrapped sample of the AffectNet training set. By leveraging both data resampling and architectural diversity, the ensemble aims to reduce variance and improve generalisation. Experiments demonstrate that the proposed ensemble achieves superior classification performance compared to individual base models, particularly in terms of robustness and Top-3 accuracy, demonstrating the effectiveness of Bagging for large-scale, imbalanced FER in unconstrained settings. This work also provides high performance benchmark using only facial cues, without augmentation or handcrafted features