Predicting at-risk students with LMS data: a comparison between Adaboost and LSTM algorithms

The prediction of students at-risk (dropout and failure) is a largely explored problem on Learning Analytics and Educational Data Mining. The present work compares the results of two different algorithms used to generate predictive models to early detect students at-risk, LSTM and Adaboost. This comparison aims to improve the performances of the models already implemented and integrated on a Moodle dashboard. For the comparison, data from a total of 122 students was collected from Moodle over four semester of an Introductory Programming course offered at Federal University of Santa Catarina (UFSC). Models were generated for each one of the 17 weeks of the semester, and their AUROC measures were then calculated and compared to evaluate the differences between LSTM and Adaboost. The results have shown that even though LSTM models presented a better performance than Adaboost, these differences were not statistically significant.