The Opportunities and Limitations of Scaling Up Sensor-Free Affect Detection

Abstract

We develop and analyze affect detectors for four affective states: confidence, excitement, frustration and interest. We utilize easy to implement self-report based “ground truth ” measurements of affect within a tutor, and model them as continuous variables that are later discretized into positive, neutral, and negative valence classifications; this distinguishes our work from detectors which model affective states as binary. We explore the opportunities and limitations of cross validation with regard to potentially distinct sample groups.