Early Detecting the At-risk Students in Online Courses Based on Their Behavior Sequences

Abstract

Online learning has developed rapidly, but the participation of learners is very low. So it is of great significance to construct a prediction model of learning results, to identify students at risk in time and accurately. We select nine online learning behaviors from one course in Moodle, take one week as the basic unit and 5 weeks as the time node of learning behavior, and the aggregate data and sequence data of the first 5 weeks, the first 10 weeks, the first 15 weeks, the first 20 weeks, the first 25 weeks, the first 30 weeks, the first 35 weeks and the first 39 weeks are formed. Eight classic machine learning methods, i.e. Logistic Regression (LR), Naive Bayes (NB), Radom Forest (RF), K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Iterative Dichotomiser3 (ID3), Classification and Regression Trees (CART), and Neural Network (NN), are used to predict the learning results in different time nodes based on aggregate data and sequence data. The experimental results show that sequence data is more effective than aggregate data to predict learning results. The prediction AUC of RF model on sequence data is 0.77 at the lowest and 0.83 at the highest, the prediction AUC of CART model on sequence data is 0.70 at the lowest and 0.83 at the highest, which are the best models of the eight classic prediction models. Then Radom Forest (RF) model, Classification and Regression Trees (CART) model, recurrent neural network (RNN) model and long short term memory (LSTM) model are used to predict learning results on sequence data; the experimental results show that long short term memory (LSTM) is a model with the highest value of AUC and stable growth based on sequence data, and it is the best model of all models for predicting learning results.