Automated Student Model Improvement

Student modeling plays a critical role in developing and improving instruction and instructional technologies. We present a technique for automated improvement of student models that leverages the DataShop repository, crowd sourcing, and a version of the Learning Factors Analysis algorithm. We demonstrate this method on eleven educational technology data sets from intelligent tutors to games in a variety of domains from math to second language learning. In at least ten of the eleven cases, the method discovers improved models based on better test-set prediction in cross validation. The improvements isolate flaws in the original student models, and we show how focused investigation of flawed parts of models leads to new insights into the student learning process and suggests specific improvements for tutor design. We also discuss the great potential for future work that substitutes alternative statistical models of learning from the EDM literature or alternative model search algorithms.