Data-Driven Student Clusters Based on Online Learning Behavior in a Flipped Classroom with an Intelligent Tutoring System

The idea of clustering students according to their online learning behavior has the potential of providing more adaptive scaffolding by the intelligent tutoring system itself or by a human teacher. With the aim of identifying groups of students who would benefit from the same intervention, in this paper, we study a set of 104 weekly behaviors observed for 26 students in a blended learning environment with AC-ware Tutor, an ontology-based intelligent tutoring system. Online learning behavior in AC-ware Tutor is described using 8 tracking variables: (i) the total number of content pages seen in the learning process; (ii) the total number of concepts seen in the learning process; (iii) the total content proficiency score gained online; (iv) the total time spent online; (v) the total number of student logins to AC-ware Tutor; (vi) the stereotype value after the initial test in AC-ware Tutor, (vii) the final stereotype value in the learning process, and (viii) the mean stereotype variability in the learning process. The previous measures are used in a four-step analysis process that includes the following elements: data preprocessing (Z-score normalization), dimensionality reduction (Principal component analysis), the clustering (K-means), and the analysis of a posttest performance on a content proficiency exam. By using the Euclidean distance in K-means clustering, we identified 4 distinct online learning behavior clusters, which we designate by the following names: Engaged Pre-knowers, Pre-knowers Non-finishers, Hard-workers, and Non-engagers. The posttest proficiency exam scores were compared among the aforementioned clusters using the Mann-Whitney U test.