Student Program Classification Using Gated Graph Attention Neural Network

Abstract

Source code mining has received increasing attention, among which code classification plays a significant role in code understanding and automatic coding. Most source code mining efforts aim at the source code of projects, which are usually large and standardized, but less for student programs. There are two differences between project codes and student programs. On the one hand, some work on project codes is based on relatively single information, which is far from enough for student programs. Because student programs are relatively small, which makes them contain less information. Consequently, it is necessary to mine as much information as possible in student programs. On the other hand, the variable or function naming and the structure of the student programs are usually irregular, as compared with the source codes of projects. To learn from student programs, we proposed a Graph Neural Network (GNN) based model, which integrates data flow and function call information to the Abstract Syntax Tree (AST), and applies an improved GNN model to the integrated graph to achieve the state-of-art student program classification accuracy. The experiment results have shown that the proposed work can classify student programs with accuracy over 97%.

Keywords

• Abstract syntax tree
• attention mechanism
• code understanding
• function call graph
• data flow
• graph neural network