Binary emotion valence classification from smartphone interactions and wristband sensors

We present a research study on a binary classification of emotions, through data gathered with the motion sensors and keystrokes of a smartphone and an equipped wristband. To this end, we developed a mobile application to induce emotions through videos and to record users’ interactions. A specific digital self-reporting system was developed from the Geneva Emotion Wheel to help participants reporting their emotions. The sensor records have been labelled according to the participants’ self-reporting and the video conditions. A method is thus proposed to process the collected time series data and to automatically classify the valence of the declared emotions using machine learning techniques. We tried a general valence classification using all emotions of all individuals and a personalised classification using a subset of emotions of a single individual. The most promising result has been obtained with a personalised model, for which we were able to achieve, on average across all participants, two-thirds of correct valence classification, using merged data from different modalities.