Reproducing Predictive Learning Analytics in CS1

Abstract

Predictive learning analytics has been widely explored in educational research to improve student retention and academic success in an introductory programming course in computer science (CS1). General-purpose and interpretable dropout predictions still pose a challenge. Our study aims to reproduce and extend the data analysis of a privacy-first student pass–fail prediction approach proposed by Van Petegem and colleagues (2022) in a different CS1 course. Using student submission and self-report data, we investigated the reproducibility of the original approach, the effect of adding self-reports to the model, and the interpretability of the model features. The results showed that the original approach for student dropout prediction could be successfully reproduced in a different course context and that adding self-report data to the prediction model improved accuracy for the first four weeks. We also identified relevant features associated with dropout in the CS1 course, such as timely submission of tasks and iterative problem solving. When analyzing student behaviour, submission data and self-report data were found to complement each other. The results highlight the importance of transparency and generalizability in learning analytics and the need for future research to identify other factors beyond self-reported aptitude measures and student behaviour that can enhance dropout prediction.