AStrId: An Interpretable Code Representation and Distance for Classifying Student Solution Strategies
(2026) IEEE International Conference on Source Code Analysis & Manipulation (SCAM) — Location: Benevento, Italy (14.September.2026)
Files
AStrId_SCAM_2026.pdf
Open Access - Adobe PDF
- 392.17 KB
Details
- Authors
- Abstract
- (en) To provide more effective feedback, instructors can benefit from automatically classifying student programs according to underlying solution strategies. We propose an automated classification approach based on AStrId, a novel abstract code representation, combined with a tree edit distance that enables structured comparison between student programs and representative strategy implementations. AStrId balances classification performance, instructor effort, and interpretability by capturing the essential structural complexity of programs while abstracting away Semantic-Preserving Variations. We evaluate our approach against two alternative representations based on JPlag and ASTs, using three training set configurations with varying labelling effort. The evaluation is conducted on 1454 programs from seven exercises in an introductory Python programming course. Our results show that AStrId achieves strong classification performance (F1-scores exceeding 0.9), even with reduced training data, thereby limiting instructor effort. At the same time, it preserves interpretability by providing insights into how and why student programs are matched to specific solution strategies.
- Affiliations
Citations
Steveny, G., Mens, K., & Nijssen, S. (2026). AStrId: An Interpretable Code Representation and Distance for Classifying Student Solution Strategies. IEEE International Conference on Source Code Analysis & Manipulation (SCAM), Benevento, Italy. https://hdl.handle.net/2078.5/279560