An Educational Psychology Inspired Approach to Student Interest Detection in Valence-Arousal Space

R. Yamamoto Ravenor, Diana Borza

Abstract: Studies on AI-based facial emotion recognition (FER) of students explore a multiplicity of algorithmic solutions, taking a categorical or a dimensional approach, but overlooking the more education-relevant and objectively measurable parameters of the latter. The “arousal” dimension aligns the degrees of emotional intensity with educational distinctions of passive-active student affect. We build on a theoretical tradition of learning where the gradual transition between these two emotional states is explained with direct reference to student interest detection. The framework proposed models student interest by passive-active emotional intensity (arousal), and only incidentally by negative-positive emotional tone (valence), attributes estimated using a resource-efficient multi-task convolutional neural network. An emotional descriptor based on a 2D histogram in the valence-arousal space is used to establish a student “baseline emotional profile” from which to run analysis on any subsequent session. This representation is explainable and graphic, allowing for simple deterministic rule-based algorithms to spot changes in arousal.

External IDs:dblp:conf/aied/RavenorB24