Are the Performance Prediction Models in MOOC General: Perspective from Big Data

Online education has become the first choice for millions of Internet users around the world to get access to knowledge and skills. Many studies aim to implement performance prediction to mark learners with high risk of poor performance in the early stage of a course. Among these researches, the employment of features concerned with learning behaviors combined with machine learning techniques proves effective to achieve decent prediction results. However, the experiments of most existing researches are based on one or two courses, making people wonder whether there is a universally applicable prediction model for multiple courses. In this paper, we investigate performance of different prediction models with the perspective of big data. We collect the background data from 28 representative courses in different subjects on a popular MOOC platform in China, and evaluate the effectiveness of the existing predict models on them. The results show that 1) Within the scope of one course, it is wise to combine different activities data for prediction, and XGBoost is most effective prediction methods in most cases;2) There does not exist a generally feasible prediction model that works for all courses.