When is a machine representation accurate enough to be learned from? When it is, what remedies to problems of educational concern does it allow to be responsibly addressed at scale? In this talk, I will present empirical works grappling with these questions in the context of higher ed, on-campus and online. Unlike the K-12 context, taxonomies are scant in higher ed and unstructured or semi-structured data exceed structured data in abundance. Connectionist (i.e., neural) approaches have been effective in learning structure from natural language and, in work presented from my lab, they are employed to learn structure from the non-linguistic data sequences that are learner pathways; in particular, millions of student enrollments at a public liberal arts university and course clickstream histories in a for-degree-credit online course on earth science. Through an AI-assistive epistemological process, meaning is made from these machine learned representations of courses and course content that are validated against domain knowledge, visualized, and reasoned about. I will describe how these representations combined with text of course content have contributed to student guidance systems, course syllabus iteration, and the potentially transformative effort to create more equitable transfer pathways between 2-year and 4-year degree granting institutions.
Zachary Pardos is an Assistant Professor at the University of California, Berkeley in the Graduate School of Education and School of Information. He directs the Computational Approaches to Human Learning research lab and teaches courses on data mining and analytics, the history of digital learning environments, and machine learning in education. His focal areas of study are knowledge representation and personalized supports leveraging big data in education.