Multi-task Learning for Stress Recognition

Different people experiencing the same stressor can have different psycho-physiological reactions. Such differences happen even when reporting the same affective state. While this mismatch causes a drop in affect recognition performance, researchers have shown that this same psycho-physiological uniqueness can be utilized to develop sensor-based biometric recognition systems. These findings open up the possibility to exploit the benefits of the psycho-physiological uniqueness observed in biometric recognition systems to improve affect machine learning models. In this paper, we explore the joint learning of a model for stress recognition and user identification to improve the performance of the first task. The joint model is learned using Multi-task Learning (MTL) neural network. We show that MTL achieves a 6.5 percentage points improvement compared to single-task models (from 86.1 to 92.6 F1-score). These results inform the design and development of stress recognition systems.