srcML-DKT: Enhancing Deep Knowledge Tracing with Robust Code Representations from srcML

Knowledge Tracing (KT) models predicting student performance in intelligent tutoring systems have been successfully deployed in several educational domains. However, their usage in open-ended programming problems poses multiple challenges due to the complexity of the programming code and a complex interplay between syntax and logic requirements embedded in code development. As a result, traditional Bayesian Knowledge Tracing (BKT) and more advanced Deep Knowledge Tracing (DKT) approaches that use binary correctness data find limited use. Code-DKT [26] is a knowledge tracing approach that uses recurrent neural networks to model learning progress leveraging information extracted from the student-generated code, incorporating abstract syntax tree (AST)-based code features, but its reliance on parsable code limits its effectiveness; unparsable submissions may constitute a substantial part of code submitted for evaluation within platforms for automated assessment of programming assignments. To overcome the ASTs limitations, we propose srcML-DKT, an extension of Code-DKT that utilizes srcML-based code representations, enabling feature extraction from both parsable and unparsable code. By capturing syntactic and structural details directly from the code text, srcML-DKT enables including all student code submissions, regardless of syntax errors. Empirical evaluations on a dataset of 610 students and six programming tasks focused on conditional statements demonstrate that srcML-DKT consistently outperforms both Code-DKT and traditional DKT models, achieving higher AUC and F1-scores across first and all attempts. These results highlight the model's ability to track student knowledge progression more accurately, in environments where trial-and-error learning is common.