Academic Performance Prediction with Machine Learning Techniques

Abstract

Nationally, the six year graduation rate for four year degree programs at universities and colleges in the United States has remained approximately 60% over the past decade. One of the main reasons for poor retention (and ultimately training) of students has been lack of proper advising and planning. Recently, there has been the prevalence of educational technologies driven by data analytics in educational environments for assisting students in selecting courses, acquiring feedback and improving learning outcomes based on past academic performance and behaviors. Grade prediction methods seek to estimate a grade that a student may achieve in a course/task that she may take in the future (e.g., next term, next assessment). Existing grade prediction methods are mainly based on matrix factorization (MF) approaches, and overlook important factors that could greatly influence student’s performance. In this thesis, I present developed several methods for performance estimation for students within a traditional brick-and-mortar university and online courses. Specifically, I model the evolution of a student’s knowledge while studying a sequence of courses within a matrix factorization framework. I provide a flexible framework that allows for incorporation of course-related and student-related factors like instructor, academic level and effort within a latent factor model. I also incorporate the influence of multiple co-taken courses within a semester along with student’s cumulative knowledge. I also present a deep learning based recommender system approach for predicting the grade a student will earn in a course that he/she plans to take in the next-term. The deep learning inspired approach provides added flexibility in learning the latent spaces in comparison to MF approaches. The proposed approach also incorporates instructor information besides student and course information. In addition, I also engineer student learning and engagement features from the server logs of students enrolled in a Massive Open Online Course (MOOCs). These features are incorporated within a Personalized Linear Multi-Regression model to predict within-class student’s performance in an online education environment. This thesis demonstrates the strengths of academic performance prediction on multiple benchmarks. Incorporating these within Early Warning Systems to identify students who are at risk of dropping out can lead to timely help from human advisors in helping students succeed within their academic programs. Accurate and timely prediction of students’ academic grades holds the promise for better student degree planning, personalized advising and providing timely feedback/interventions to ensure that students stay on track in their chosen degree program and graduate on time.