Affective state recognition in online learning: A deep learning approach with data augmentation

Academic affective states are crucial indicators of student engagement and participation. With the rise of digital education, understanding students’ emotional and interest levels has become more challenging. In this study, we utilise the DAiSEE video dataset and present a streamlined spatiotemporal framework that combines a customised CNN-BiLSTM architecture with time-distributed processing, and experiment with facial cropping, data augmentation, and alternative architectures. Despite our model’s simplicity, when paired with carefully chosen data augmentation techniques such as horizontal flipping and adaptive gamma brightness correction, the model outperforms larger architectures while achieving significantly lower inference latency. We show that body posture and contextual cues offer vital information beyond facial expressions through ablation studies comparing full-frame versus face-only inputs. With the right data augmentation strategy, even simple models can achieve state-of-the-art performance in educational environments.