Philip I. S. Lei

In recent years, Massive Open Online Course (MOOC) has greatly changed the way the world learns. MOOC platforms (MOOCs) offer free online courses for everyone, but the high dropout rate on MOOCs is a serious problem, so early prediction of students with dropout intentions is useful to reduce the dropout rate by taking suitable intervention. Because MOOC is an online course, the system can easily collect the users’ learning logs. Using data mining techniques on the user learning log to do prediction, this study proposed a method to extract some useful features from the user behaviour and offer a lightweight method based on the Light Gradient Boosting Machine (LightGBM) with the Optuna tuning method to predict the probability that a user would drop out of a course in next 10 days. This study also explored the effect of lower feature granularity by splitting in record period into three stages instead of splitting into weeks, which used fewer learning records of the public Knowledge Discovery and Data Mining (KDD) MOOC dataset to bring less workload than other works. The proposed method requires fewer features than previous works and thus can speed up the training time while being more scalable with a large number of users and learning activities in MOOC. The experiment results showed that the performance is similar to or higher than the related previous studies, with an AUCROC score of 89.12%, AUCPR score of 96.04% and F1-score of 92.12%. This study also examines the effect on dropout prediction accuracy when training data is limited to one- and two-thirds of the original duration and finds that comparable performance can still be achieved.

Knowledge tracing (KT) is a research field of growing importance in technology enhanced education. It models students’ mastery levels of skills and predicts their performance in question answering through an Intelligent Tutoring System (ITS). Traditionally, Bayesian methods like Bayesian Knowledge Tracing (BKT) and Weakest Knowledge Tracing (WKT) apply Hidden Markov Model to model the latent learning process of individual skills. With the advent of recent advances in Deep Learning (DL), DL based KT methods achieve better prediction performance by exploiting the temporal dependencies between consecutive exercises on different skills. However, Bayesian methods are still valuable with a simpler, more intuitive and interpretable model of learning. This paper proposes a new Bayesian method called Corrigible Knowledge Tracing (CKT) which assumes students can learn from mistakes in answering multi-skill questions. In addition, the combinations of high occurrence skills are considered, so that integrated abilities are more fully recognised. In evaluation, the harmonic mean is suggested for combining the predicted probabilities, and accuracy is chosen to be the performance metric. The proposed method is compared with both Bayesian methods and DL approaches by using the conventional question-solving records from a modern large-scale educational data mining dataset called EdNet. Experiment results show that the performance of CKT is on par with state-of-the-art DL approaches. The overall improvement from WKT to the proposed method is substantial.