Reproduce “Heterogeneous Directed Hypergraph Neural Network over abstract syntax tree (AST) for Code Classification” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Guang Yang, Tiancheng Jin, Liang Dou
HDHGN represents program ASTs as heterogeneous, directed hypergraphs in which each typed AST field forms a hyperedge connecting multiple child nodes to their parent. On Project CodeNet classification benchmarks, it reaches 97.87% accuracy on Python800 and 96.42% on Java250, outperforming the strongest reported baseline by 2.88 and 2.47 percentage points, respectively.
Key insights
- Modeling all children of an AST field within one hyperedge preserves higher-order structural information that ordinary pairwise graphs discard.
- Node heterogeneity, field-specific edge types, and directed parent-child roles each contribute to the learned representation rather than serving only as implementation details.
- Ablations show that removing hyperedges causes the largest performance drop, followed by removing heterogeneous information and then edge direction.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.