Predictive Models of Student Learning in K–16

Co-sponsored by the Graduate School of Education and the School of Information

The volume and detail of data produced by current educational technologies in K–16 present an exciting opportunity to gain new insights into learning and personalization. In this talk I will show how probabilistic graphical models of student learning, with roots in cognitive theory, have served as an effective platform to study learning phenomenon and to increase predictive accuracy within virtual learning environments. Probabilistic graphical models allow for a blend of machine learning and domain expertise and are well suited to capture the temporal aspects of student data. The models were used to improve the accuracy of student knowledge assessment and performance prediction by taking into account each student’s individual prior knowledge and learning rate. This approach proved effective in the 2010 KDD Cup competition on educational data mining. I will also show how these same models have been posed to measure the efficacy of pedagogical decisions made by the learning environment, such as the selection of items and interventions. The validity of the model’s conclusions about effective pedagogical practices, inferred from log data, is checked against the conclusions of A/B studies with pre and post-tests. The work presented was conducted using data from the Cognitive Tutors for Algebra and Geometry, the ASSISTments Platform for K–12 math and, recently, a Massive Open-access Online Course (MOOC) on the edX platform.