Predicting Tomorrow’s Mood, Health, and Stress Level using Personalized Multitask Learning and Domain Adaptation

Abstract

Predicting a person’s mood tomorrow, from data collected unobtrusively using wearable sensors and smartphones, could have a number of beneficial clinical applications; however, this prediction is an extremely challenging problem. Past approaches often lack the accurate and reliable performance necessary for real-world applications. We posit that this is due to the inability of traditional, one-size-fits-all machine learning models to account for individual differences. To overcome this, we treat predicting tomorrow’s mood for a single person as one task, or problem domain. We then adopt Multitask Learning (MTL) and Domain Adaptation (DA) approaches to learn a model which is customized for each person, while still being able to benefit from data across the population. Empirical results on real-world, continuous monitoring data show that the new personalized models — a MTL deep neural network, and a Gaussian Process with DA — both significantly outperform their generic counterparts, providing substantial performance enhancements in automatic prediction of continuous levels of tomorrow’s reported mood, stress, and physical health based on data through today