Unsupervised Discovery of Student Strategies

Unsupervised learning algorithms can discover models of student behavior without any initial work by domain experts, but they also tend to produce complicated, uninterpretable models that may not predict student learning. We propose a simple, unsupervised clustering algorithm for hidden Markov models that can discover student learning tactics while incorporating student-level outcome data, constraining the results to interpretable models that also predict student learning. This approach is robust, domain-independent, and does not require domain experts. The models have tecst-set correlations with learning gain as high as 0.5 and the findings suggest possible improvements to the scaffolding used by many software tutors.