Investigating Self-supervised Learning for Predicting Stress and Stressors from Passive Sensing

The application of machine learning (ML) techniques for well-being tasks has grown in popularity due to the abundance of passively-sensed data generated by devices. However, the performance of ML models are often limited by the cost associated with obtaining ground truth labels and the variability of well-being annotations. Self-supervised representations learned from large-scale unlabeled datasets have been shown to accelerate the training process, with subsequent fine-tuning to the specific downstream tasks with a relatively small set of annotations. In this paper, we investigate the potential and effectiveness of self-supervised pre-training for well-being tasks, specifically predicting both workplace daily stress as well as the most impactful stressors. Through a series of experiments, we find that self-supervised methods are effective when predicting on unseen users, relative to supervised baselines. Scaling both data size and encoder depth, we observe the superior performance obtained by self-supervised methods, further showcasing their utility for well-being applications. In addition, we present future research directions and insights for applying self-supervised representation learning on well-being tasks.