Stress detection on smartphone data with a machine learning approach based on Mahalanobis distance-based outlier finding and ReliefF feature selection

Stress is beneficial when a person is focused, awake and alert. However, exposure to high doses of stress harms a person's health. For this reason, it is important to detect stress and begin relief as soon as possible. In this study, soft keyboard typing behaviors with touchscreen panel, gravity, linear acceleration, and gyroscope data obtained from smartphones were examined. It was observed that there was a correlation between the results obtained and typing behaviors and the stress levels of individuals. In this context, an expanded data set was created. In order to detect stress with higher accuracy, a Mahalanobis distance-based outlier detection approach was applied. Subsequently, a structure was created by combining the ReliefF feature selection method and machine learning techniques to identify efficient features and perform classification. The results obtained by cleaning outlier data showed that the created structures achieved success with high accuracy. In addition, outlier detection and cleaning increased the classification success by 1.77 points.