Scholarlink

<p>Automated assessment systems such as online judge systems constitute the primary infrastructure for large-scale, deliberate programming practice. Although these systems can verify functional correctness of a submission, they remain agnostic to the solution strategies adopted by students. Previous studies have predominantly relied on unsupervised clustering methods to group solutions. However, these methods fall short of pedagogical demands due to their limited generalizability across tasks and neglect of fine-grained strategic patterns. To overcome these limitations, we introduce SchemaNet, a high-accuracy supervised model that detects student solution strategies via multi-view features, especially code schemas extracted at sub-problem granularity by our proposed SchemaMiner framework. These schemas are then fused with static (AST-based) and dynamic (execution trace-based) representations to capture a comprehensive view of student problem-solving approaches. We have validated our approach on a newly constructed, manually annotated dataset of 1,612 Python solutions across seven diverse introductory problems. Experimental results demonstrate that SchemaNet achieves a state-of-the-art average accuracy of 94.77%, significantly outperforming both alternative classification baselines and clustering models. This high level of accuracy confirms the superiority of the supervised approach and indicates that the model is sufficiently reliable for practical deployment in real-world educational settings. The fully annotated dataset and implementation code are made publicly available to support further research.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

From clustering to classification: mining code schemas for accurate identification of student solution strategies

  • 期刊论文

  • 发表日期：

    2026年7月31日

查看全文
  • Ying Jiang,
  • Qing Zhou,
  • Dayu Wu,
  • Subo Zhao,
  • Cifa Lin

摘要

Automated assessment systems such as online judge systems constitute the primary infrastructure for large-scale, deliberate programming practice. Although these systems can verify functional correctness of a submission, they remain agnostic to the solution strategies adopted by students. Previous studies have predominantly relied on unsupervised clustering methods to group solutions. However, these methods fall short of pedagogical demands due to their limited generalizability across tasks and neglect of fine-grained strategic patterns. To overcome these limitations, we introduce SchemaNet, a high-accuracy supervised model that detects student solution strategies via multi-view features, especially code schemas extracted at sub-problem granularity by our proposed SchemaMiner framework. These schemas are then fused with static (AST-based) and dynamic (execution trace-based) representations to capture a comprehensive view of student problem-solving approaches. We have validated our approach on a newly constructed, manually annotated dataset of 1,612 Python solutions across seven diverse introductory problems. Experimental results demonstrate that SchemaNet achieves a state-of-the-art average accuracy of 94.77%, significantly outperforming both alternative classification baselines and clustering models. This high level of accuracy confirms the superiority of the supervised approach and indicates that the model is sufficiently reliable for practical deployment in real-world educational settings. The fully annotated dataset and implementation code are made publicly available to support further research.