Learning Models for Student Performance Prediction

Predicting student performance supports educational decision-making by allowing directives and teachers to detect students in special situations (e.g. students at risk of failing a course or dropping out of school) and manage these in a timely manner. The problem we address consists of grade prediction for the courses of a given academic period. We propose to learn a predictive model for each course. Two cases can be distinguished: historical grades are unavailable for prediction (first semester) and historical grades are available. For the first case, features that include selection test scores, socioeconomic information, and middle school the student comes from are proposed. For the second case, features that include past grades from similar courses are proposed. To test our approach, we gathered data from a Mexican public high school (three generations, 2,000 students, four semesters, and 24 courses). Our results indicate that features such as numerical ability, family, motivation, and social sciences are relevant for prediction without historical grades, while grades from the immediate previous semester are relevant for prediction with historical grades. Additionally, support vector machines and linear regression are suitable techniques for tackling grade prediction.