Emotion detection from smartphone keyboard interactions: role of temporal vs spectral features.

Keystroke or typing dynamics represent two key facets - rhythm corresponds to spectral-domain characteristics and timing corresponds to time-domain behavior, which are created when a person types. The presence of inherent time-domain and frequency-domain characteristics in smartphone keyboard interactions motivate us to perform a comparative analysis of time-domain and frequencydomain features for emotion detection. We design, and develop an Android-based data collection application, which collects keyboard interaction logs and emotion self-reports (happy, sad, stressed, relaxed) from 18 subjects in a 3-week in-the-wild study. For the time-domain analysis, we extract a set of time-domain features and construct Random Forest-based personalized model; whereas for the spectral-domain analysis, first transform the interaction details into frequency-domain using DFT (Discrete Fourier Transform) and then extract a set of spectral-domain features to construct a personalized model for emotion detection. The empirical analysis from the study reveals that the time-domain models return superior classification performance (average AUCROC 72%) than the frequency-domain models (average 67%). It also signifies the importance of several time-domain and frequency-domain features as a strong discriminator of emotion states.