Meta-Reasoning Algorithm for Improving Analysis of Student Interactions with Learning Objects using Supervised Learning

Supervised learning (SL) systems have been used to automatically learn models for analysis of learning object (LO) data. However, SL systems have trouble accommodating data from multiple distributions and “troublesome” data that contains irrelevant features or noise—all of which are relatively common in highly diverse LO data. The solution is to break up the available data into separate areas and then take steps to improve models on areas containing troublesome data. Unfortunately, finding these areas in the first place is a far from trivial task that balances finding a single distribution with having sufficient data to support meaningful analysis. Therefore, we propose a BoU metareasoning (MR) algorithm that first uses semi-supervised clustering to find compact clusters with multiple labels that each support meaningful analyses. After clustering, our BoU MR algorithm learns a separate model on each such cluster. Finally, our BoU MR algorithm uses feature selection (FS) and noise correction (NC) algorithms to improve models on clusters containing troublesome data. Our experiments, using three datasets containing over 5000 sessions of student interactions with LOs, show that multiple models from BoU MR achieve more accurate analyses than a single model. Further, FS and NC algorithms are more effective at improving multiple models than a single model.