Reproduce “Self-Attention Graph Pooling” 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

Junhyun Lee, Inyeop Lee, Jaewoo Kang

Self-Attention Graph Pooling (SAGPool) introduces a hierarchical pooling mechanism that leverages graph convolution to compute self-attention scores for topology-aware node selection. This approach consistently outperforms existing hierarchical graph pooling methods on benchmark graph classification datasets, demonstrating superior efficiency and scalability.

Key insights

  • Explicitly considering graph topology during attention-based node selection for pooling is critical for achieving improved performance in graph classification tasks.
  • Self-attention, when computed through graph convolution, provides an effective and learnable mechanism to dynamically identify and retain the most salient nodes for hierarchical graph coarsening.
  • Efficient and scalable hierarchical graph pooling can be achieved without sacrificing performance, addressing limitations of prior methods regarding complexity and parameter overhead.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.