Use of Behavior Dynamics to Improve Early Detection of At-risk Students in Online Courses

Abstract

Online learning has developed rapidly, but the success rate is very low. Hence, it is of great significance to construct a learning result predicting model, and to quickly and accurately identify students at risk of failing their course. In order to mine the dynamic features of learning behaviors and use them to improve the accuracy of detection of at-risk students, we propose a long-short term memory (LSTM) network based approach to identify at-risk students. To validate the performance of this approach, we first extracted the behavior data of one course from a public dataset, and generate two types of datasets, the aggregated datasets and the sequential datasets. After that, we used eight classic machine learning methods to train predicting model on these datasets and explored whether the models trained on sequential datasets are more accurate than the models trained on aggregated datasets. The results show that the models trained on sequential datasets are more accurate when naïve Bayes, Classification and Regression Tree, Random Forest (RF), Iterative Dichotomiser 3 and Multilayer Perception are used. Finally, we used the LSTM to train predicting models on sequential datasets, and compared them with the best models trained by RF. The results show that the models trained by the LSTM are more accurate, which proves the effectiveness of the proposed approach at certain extent.