Learning Effective Models of Emotions from Physiological Signals: The Seven Principles

Learning effective models from emotion-elicited physiological responses for the classification and description of emotions is increasingly required to derive accurate analysis from affective interactions. Despite the relevance of this task, there is still lacking an integrative view of existing contributions. Additionally, there is no agreement on how to deal with the differences of physiological responses across individuals, and on how to learn from flexible sequential behavior and subtle but meaningful spontaneous variations of the signals. In this work, we rely on empirical evidence to define seven principles for a robust mining physiological signals to recognize and characterize affective states. These principles compose a coherent and complete roadmap for the development of new methods for the analysis of physiological signals. In particular, these principles address the current over-emphasis on feature-based models by including critical generative views derived from different streams of research, including multivariate data analysis and temporal data mining. Additionally, we explore how to use background knowledge related with the experimental setting and psychophysiological profiles from users to shape the learning of emotion-centered models. A methodology that integrates these principles is proposed and validated using signals collected during human-to-human and human-to-robot affective interactions.