Course-specific markovian models for grade prediction

Abstract

Over the past 15 years, the average six-year graduation rates for colleges and universities across the Unites States have remained stable at around 60%. This vehemently impacts society in terms of workforce development, national productivity and economic activity. Educational early-warning systems have been identified as an important approach to tackle this problem. The key to these systems are accurate grade prediction algorithms. In this paper we propose application of markovian models for the problem of predicting next-term student performance. Traditional approaches predict student’s grade in a course by using a subset of prior courses and content features. However, these models ignore the dynamic evolution of student’s knowledge states, which is a strong influence on student’s learning and performance. We developed course-specific Hidden Markov Models and Hidden Semi-markov Models for the problem of next-term grade prediction. Our experimental results on datasets from a large public university show that the proposed approaches outperform prior state-of-the-art methods. We show by a case study the application of these methods for early identification of at-risk students.