Knowledge Tracing Within Single Programming Exercise Using Process Data

Abstract

Knowledge tracing is a core technology in many intelligent learning systems. In this paper, we propose a novel knowledge tracing method that predicts learner’s knowledge state within a single programming exercise. Given a programming task, a student’s intermediate solution is represented by an abstract syntax tree and evaluated by computing its tree edit distance to the best solution. With the measure of solution quality, the learning trajectory of each student can be encoded as a real-valued sequence. Using the mean value of the sequence as a primary feature, we developed a logistic regression model to predict students’ knowledge state. We compared our method with three popular models on a large-scale dataset collected from a classic block-based programming task. The experimental results suggest that the proposed method that captures features derived from student's problem-solving processes can significantly improve the prediction performance.