Reproduce “Agent-based Graph Neural Networks” 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
Karolis Martinkus, Pál András Papp, Benedikt Schesch, Roger Wattenhofer
Researchers at ETH Zurich and Huawei developed AgentNet, an architecture that uses multiple neural agents to learn intelligent graph exploration for graph-level tasks. This approach achieves expressiveness beyond 2-Weisfeiler-Lehman (WL) and demonstrates sublinear computational complexity, maintaining competitive accuracy on diverse real-world benchmarks while efficiently processing larger graphs.
Key insights
- The paper demonstrates that fully differentiable, learned graph traversals by neural agents can overcome the expressiveness limitations of traditional GNNs, distinguishing graphs beyond 2-WL.
- AgentNet inherently offers the potential for sublinear computational complexity for graph-level tasks, as agents can learn to identify class-defining substructures by exploring only a fraction of the graph.
- The use of multiple, independently operating agents enhances expressiveness and efficiency, allowing for parallelized and focused exploration to detect specific graph properties more robustly.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.