Meta-Learning for Knowledge Tracing in Student Coding Data

Anonymous

Keywords: Knowledge Tracing, Meta-Learning, Deep Learning, CS Education, Natural Language Processing, LSTMs, Memory Augmented Neural Networks

TL;DR: We employ a meta-learning approach to solve the problem of predicting whether or not a student will struggle on future coding problems given the student's performance on past coding problems.

Abstract: Predicting whether a student will be able to solve new coding problems given past performance is an important but challenging task. Employing a traditional supervised machine learning approach to tackle this classification problem is intractable because of the large amount of data needed per student to make reasonable predictions. In this paper, we employ a meta-learning approach to solve this issue, where each task corresponds to predicting a single student's coding performance given data on that student's past performance; this allows us to exploit the benefits of shared structure that individual student's submission data would have. Our best model, a Memory Augmented Neural Network (MANN) architecture which concatenates submission history embeddings to handcrafted submission features and embeddings representing coding problem concepts, achieves an accuracy of 90.5 percent and F1 score of 90.9 percent on this problem. We find that adding submission history embeddings and coding concept embeddings to our baseline model increases the accuracy by 4 percent and F1 score by 3 percent.