Abstract:
Stress is one of the most important problems of today. Although it seems to be a part of modern human life, it is known to cause serious health problems. Many researchers from different disciplines have been working on this subject, which has personal and social effects, for many years. Psychologists, behavioral scientists, and psychiatrists continue their research in the clinical setting. However, when the stress factor is considered as a part of daily life, clinical environments or controlled experimental areas may be insufficient in terms of stress classification. Thanks to developing sensor technologies, wearable devices, and machine learning methods, stress classification has become an area of interest for computer scientists. Although developments in wearable sensors, ubiquitous computing, and machine learning continue, they bring new challenges to this field. The data labeling burden is one of these challenges. It requires significant effort and resources to have the subjects who have stress problems fill out questionnaires periodically in their daily life and to synchronize the physiological data with the questionnaire results. Being aware of this labeling burden, we aimed to find a new solution by using a less amount of labeled data from the multi-sensor physiological dataset that we collect in daily life. For this reason, this thesis focuses on what will be the performance of a system using a less amount of labeled data and semi-supervised learning techniques.