Learning from, with and Without the Interdependencies of Valence-Arousal-Dominance and Their Connection with Basic Emotions

The majority of approaches to affective computing that deal with Valence, Arousal and Dominance (VAD) ratings neglect their interdependencies. Although initially postulated as independent dimensions, several behavioral studies have shown that they share mutual information. In this paper, we show that considering the shared variance between the three dimensions is non-negligible for several applications. Therefore, we use inter-dimensional regression models to extract residual VAD ratings and show how they can help illuminate and align to nomothetic differences between demographic groups and with mapping VAD ratings to basic emotions.