Frustratingly Easy Personalization for Real-time Affect Interpretation of Facial Expression

In recent years, researchers have developed technology to analyze human facial expressions and other affective data at very high time resolution. This technology is enabling researchers to develop and study interactive robots that are increasingly sensitive to their human interaction partners' affective states. However, typical interaction planning models and algorithms operate on timescales that are frequently orders of magnitude larger than the timescales at which real-time affect data is sensed. To bridge this gap between the scales of sensor data collection and interaction modeling, affective data must be aggregated and interpreted over longer timescales. In this paper we clarify and formalize the computational task of affect interpretation in the context of an interactive educational game played by a human and a robot, during which facial expression data is sensed, interpreted, and used to predict the interaction partner's gameplay behavior. We compare different techniques for affect interpretation, used to generate sets of affective labels for an interactive modeling and inference task, and evaluate how the labels generated by each interpretation technique impact model training and inference. We show that incorporating a simple method of personalization into the affect interpretation process - dynamically calculating and applying a personalized threshold for determining affect feature labels over time - leads to a significant improvement in the quality of inference, comparable to performance gains from other data pre-processing steps such as smoothing data via median filter. We discuss the implications of these findings for future development of affect-aware interactive robots and propose guidelines for the use of affect interpretation methods in interactive scenarios.