Using Bayesian Networks to Predict Student Performance

Bayesian networks trained on students' core physics course performance data were used to analyze student progression through the curriculum and predict student performance in a set of upper-level undergraduate courses. An expert-elicited method of developing conditional dependency structures is explored and compared to standard Bayesian network structure learning algorithms. The course grade classification performance of a set of Bayesian network models is examined for modern physics, electricity and magnetism, and quantum mechanics courses. Adjustments to modeling procedures are explored in consideration of low record counts in the course performance datasets. Benefits of Bayesian networks, such as ease of interpretation and conditional probability querying, are discussed.