Use of Unsupervised Clustering to Characterize Learner Behaviors and Affective States while Using an Intelligent Tutoring System

Document Type


Publication Date



This paper presents results from a preliminary analysis of interaction and human observation data gathered from students using an Aplusix, an intelligent tutoring system for algebra. Towards the development of automatic detectors of behavior and affect, this study tried to determine whether it was possible to identify distinct groups of students based on interaction logs alone. Using unsupervised clustering, we were able to identify that student behaviors within the software cluster into two categories, Clusters 0 and 1, associated with differing higher-level behaviors and affective states. Cluster 0 tended to reflect more collaborative work, whereas Cluster 1 reflected more solitary work. Cluster 1 students tended to exhibit more flow, suggesting that students in flow tend to work in a more individual fashion. An examination of the keystrokes used by each group showed that Cluster 0 used the arrow keys and cursor keys significantly more than Cluster 1. The Cluster 1, on the other hand, tended to type more mathematical operators or use the duplicate command more frequently than Cluster 0. This implies that frequent use of mathematical operators and frequent duplication of the problem may be evidence of flow within Aplusix.