Reproduce “Simplifying Graph Convolutional 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
Felix Wu, Tianyi Zhang, Amauri Holanda de Souza Jr., Christopher Fifty, Tao Yu, Kilian Q. Weinberger
Researchers from Cornell University's Kilian Q. Weinberger group developed Simple Graph Convolution (SGC), a linear model derived from Graph Convolutional Networks (GCNs) by removing non-linearities and collapsing weight matrices. SGC achieves competitive or superior performance across various graph learning tasks while being up to two orders of magnitude faster and significantly more memory efficient than complex GCN variants.
Key insights
- The core effectiveness of GCNs stems primarily from their graph propagation mechanism, which smooths features across the graph, rather than from intermediate non-linear feature transformations.
- SGC acts as a fixed low-pass filter in the graph spectral domain, where features are smoothed over K-hop neighborhoods.
- The 'renormalization trick' (adding self-loops to the adjacency matrix) is crucial as it theoretically shrinks the spectrum of the normalized graph Laplacian, ensuring that graph propagation acts as a stable low-pass filter and preventing issues like exploding filter coefficients.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.