MOOC Performance Prediction by Deep Learning from Raw Clickstream Data

Abstract

Student performance prediction is a challenging problem in online education. One of the key issues relating to the quality Massive Open Online Courses (MOOC) teaching is the issue of how to foretell student performance in the future during the initial phases of education. While the fame of MOOCs has been rapidly increasing, there is a growing interest in scalable automated support technologies for student learning. Researchers have implemented numerous different Machine Learning algorithms in order to find suitable solutions to this problem. The main concept was to manually design features through cumulating daily, weekly or monthly user log data and use standard Machine Learners, like SVM, LOGREG or MLP. Deep learning algorithms could give us new opportunities, as we can apply them directly on raw input data, and we could spare the most time-consuming process of feature engineering. Based on our extensive literature survey, recent deep learning publications on MOOC sequences are based on cumulated data, i.e. on fine-engineered features. The main contribution of this paper is using raw log-line-level data as our input without any feature engineering and Recurrent Neural Networks (RNN) to predict student performance at the end of the MOOC course. We used the Stanford Lagunita’s dataset, consisting of log-level data of 130000 students and compared the RNN model based on raw data to standard classifiers using hand-crafted commulated features. The experimental results presented in this paper indicate the RNN’s dominance given its dependably superior performance as compared with the standard method. As far as we know, this will be the first work to use deep learning to predict student performance from raw log-line level students’ clickstream sequences in an online course.