Proposing Design Recommendations for an Intelligent Recommender System Logging Stress

The connection between stress and smartphone usage behavior has been investigated extensively. While the prediction results using machine learning are encouraging, the challenge of how to cope with data loss remains. Addressing this problem, we propose an Intelligent Recommender System for logging stress based on adding a subjective user data-based validation to predictions made by intelligent algorithms. In a user study involving 731 daily stress self-reports from 30 participants we found discrepancies between subjective and smartphone usage data, i.e. battery, call information, or network usage. Despite the good prediction accuracy of 65% using a Random Forest classifier, combining both information would be beneficial for avoiding data and improving prediction accuracy. For realizing such a system (i.e., a mobile application), we propose three design recommendations, based on the capabilities of frequently used machine learning classifiers, enabling users to annotate their daily stress levels with a predict-and-validate methodology.