Authentic Emotion Classification: A Machine Learning Study on Therapy Clips

The recognition and classification of emotions play a crucial role in human life. In therapeutic contexts, this plays a crucial role in improving the therapeutic process by enabling customized interventions. Traditional methods often rely on subjective interpretation, which can vary significantly among therapists, underscoring the necessity for reliable and objective approaches. This study aims to develop computational models for predicting emotions from video material of real therapy sessions, building on Ekman’s basic emotions framework and extending it with additional emotional dimensions. The dataset consists of authentic therapy session clips. Psychology students, trained in emotion labeling, provided multidimensional labels for these clips, forming the basis for our model training. We employed both traditional machine learning and neural network techniques to predict these labels. Findings indicate that our classification framework achieves emotion recognition accuracy above chance levels, offering a valuable tool for therapists and researchers. However, model accuracy is limited by the inherent variability and subjectivity of human labeling. This paper contributes to the evolving field of automated emotion classification, providing insights into model performance and potential applications in therapeutic settings.