Towards Better Affect Detectors: Detecting Changes Rather Than States

Affect detection in educational systems has a promising future to help develop intervention strategies for improving student engagement. To improve the scalability, sensor-free affect detection that assesses students’ affective states solely based on the interaction data between students and computer-based learning platforms has gained more and more attention. In this paper, we present our efforts to build our affect detectors to assess the affect changes instead of affect states. First, we developed an affect-change model to represent the transitions between the four affect states; boredom, frustration, confusion and engagement concentration with ASSISTments dataset. We then reorganized and relabeled the dataset to develop the affect-change detector. The data science platform (e.g., RapidMiner) was adopted to train and evaluate the detectors. The result showed significant improvements over previously reported models.