Logistic Regression in a Dynamic Bayes Net Models Multiple Subskills Better!

A single student step in an intelligent tutor may involve multiple subskills. Conventional approaches either sidestep this problem, model the step as using only its least known subskill, or treat the subskills as necessary and probabilistically independent. In contrast, we use logistic regression in a Dynamic Bayes Net (LR-DBN) to trace the multiple subskills. We compare these three types of models on a published data set from a cognitive tutor. LR-DBN fits the data significantly better, with only half as many prediction errors on unseen data.