Predicting Student Performance on Digital Learning Platforms Using Artificial Intelligence

Abstract

Digital learning platforms record detailed traces of student activity. This study evaluates artificial intelligence methods that use these traces to identify students at risk of failing or withdrawing early enough for support to help.

The Open University Learning Analytics Dataset, with demographic records, assessment results and 10.6 million daily click records for 32,593 course registrations, was used. At-risk status, defined as a final result of Fail or Withdrawn, was predicted at weeks 0 to 12 of each course from demographic, engagement and assessment features. Logistic regression, random forest, multilayer perceptron, XGBoost and LightGBM were trained on the 2013 course presentations and tested on the 2014 presentations.

With demographic data alone, the area under the ROC curve (AUC) was 0.68. It increased to 0.76 at week 4, 0.78 at week 6 and 0.82 at week 12. At week 6, XGBoost identified 63.3% of at-risk students with a precision of 67.9%. Differences between models were small, and logistic regression remained competitive. The mean assessment score, missing assessments and recent platform activity were the strongest predictors.

The study shows week by week when a useful warning becomes possible, using a protocol that trains on past cohorts and tests on a new one. A group analysis revealed lower accuracy and more false alarms for female students, which calls for fairness audits before such systems are deployed.