Investigating Keystroke Dynamics and Their Relevance for Real-Time Emotion Recognition

Abstract

There is strong evidence that emotional states affect the Human’s performance and decision making. Therefore, understanding Human emotions has become of great concern in the field of Human Computer Interaction (HCI). One way to online emotion recognition is through Keystroke Dynamics. It addresses the drawbacks of current methods which are intrusive and not user-friendly, expensive to implement, and neither realistic nor applicable in a real-world context. The keystroke dynamics approach focuses on analyzing the particular way a person types on a keyboard. In our research work, we start by developing a web application (EmoSurv) in order to collect the data and build a dataset. We generate datasets for free-text and fixed-text entries. These datasets are labeled with emotional states of the participants (Angry, Happy, Sad, Calm, and Neutral state). The obtained datasets are used for training and building models using machine learning algorithms. Outstanding accuracy rates are obtained reaching 93.922% and Kappa equal to 0.9197 using Random Committee algorithm. We finally provide a set of recommendations for future experimentation by comparing the different models generated.