Learning the Features Used To Decide How to Teach

As a step towards scaling personalized instruction, we seek to automatically identify the key features of the interactive learning process teachers use to select the next activity when teaching a single student. Such features could both inform computational student models designed to facilitate instructional decisions, and help enable automated self-improving teaching systems that leverage this identified feature set. We present preliminary results that a very small set of features is almost as good as a much larger set of features at predicting human tutor decisions when teaching students about histograms.