Simplifying Graph Convolutional Networks · CiteArk

Simplifying Graph Convolutional Networks

Public

Authors:Felix Wu, Tianyi Zhang, Amauri Holanda de Souza Jr., Christopher Fifty, Tao Yu, Kilian Q. Weinberger

ICML 2019 arXiv 2019 Machine Learning (cs.LG)Machine Learning (stat.ML)

Research summary

Based on the paper

This paper derives Simple Graph Convolution (SGC) by removing intermediate nonlinearities from a graph convolutional network and collapsing its weight matrices into one linear classifier. The resulting method applies a fixed, parameter-free graph propagation filter to node features before multinomial logistic regression. A spectral analysis relates propagation with self-loops to low-pass filtering and proves that self-loops shrink the normalized-Laplacian spectrum. Experiments compare SGC with graph neural-network baselines on citation and social networks and adapt it to text classification, geolocation, relation extraction, zero-shot image classification, graph classification, and molecular prediction, emphasizing accuracy, training time, stability, and known failure cases.

What this research means

Research interpretation

The work shows that repeated graph smoothing, rather than nonlinear transformations at every layer, can explain much of GCN performance on several benchmarks. SGC is therefore a useful simple and fast baseline. The paper does not claim universal equivalence: it reports a small loss on MR, substantially worse QM8 error than more expressive graph models, and graph-classification performance far behind GIN.

SGC Reproduction: Two Close, One Short

A ten-run reproduction of three public citation-network results found Cora and Pubmed numerically close to the paper, while Citeseer depended on a source-conflicted epoch setting. The primary comparison follows the paper-stated 100 epochs; the README’s 150 epochs are shown only as a sensitivity check.

Reproduction Gaps and Epoch Ambiguity

Relative to the paper values, the ten-run means were 0.05 percentage points higher on Cora, 0.06 higher on Pubmed, and 0.70 lower on Citeseer under the paper-stated 100 epochs. Citeseer's 150-epoch sensitivity mean rose to 71.85%, but this arm arose from a paper–README condition conflict and cannot replace the primary comparison by outcome; Pubmed's numerical evidence also remains inconclusive because independent assessment was unavailable.

View values and explanation

Relative to each paper value, the ten-run mean differences are Cora +0.05, Citeseer −0.70, and Pubmed +0.06 percentage points. All ten Citeseer runs in the 100-epoch primary arm were 71.2%; the 150-epoch sensitivity runs ranged from 71.8% to 71.9% and averaged 71.85%, while the paper value was 71.9%.

Cora | ten-run mean test accuracy: 81.05
Cora | paper test accuracy: 81
Citeseer | 100-epoch ten-run mean test accuracy: 71.2
Citeseer | paper test accuracy: 71.9
Pubmed | ten-run mean test accuracy: 78.96
Pubmed | paper test accuracy: 78.9
Citeseer | 150-epoch sensitivity mean: 71.85

Extensions and report history

Extension views

Repository details