1506 - Using Machine Learning to Identify At-Risk Students in Physics Classes

Abstract

Machine learning algorithms represent an exciting new class of quantitative methods to understand physics classes and students. Recent work has applied these algorithms to understand physics major retention to degree and the risk factors influencing success in introductory physics. This talk will explore the application of these algorithms to identify students most at risk of failure in introductory calculus-based physics. These predictions are complicated by the very unbalanced nature of the samples; most students do not fail physics. Machine learning algorithms will be explored in depth including decision trees and random forests. These will be applied to understand student risk both using institutional variables such as college GPA and within class variables such as homework grades with the goal of making accurate predictions early in the class. Limitations and data requirements for these algorithms as well as their accuracy for underrepresented populations will also be discussed.