Accurate Prediction and Analysis of College Students" Performance from Online Learning Behavior Data

Abstract

In order to improve accuracy in the prediction of college students" performance, a collection of students" online learning behaviors is used as input for bidirectional long short-term memory with a self-attentive mechanism to build a performance prediction model. The model is compared with K-means and LadFG algorithms in simulation experiments. The results classify students" online learning behaviors into four types (stagnant, focused, catch-up, and planned) with weighted accuracy at 0.886 and a weighted F1-score of 0.882. In the ablation experiment, the prediction model before ablation produced weighted accuracy of 0.908 and a weighted F1-score of 0.904, whereas weighted accuracy after ablation was 0.834 and the weighted F1-score was 0.835.