Simplifying Graph Convolutional Networks

With no regularization and L-BFGS, the Reddit SGC training is reported to converge in two steps; evaluation is inductive, training on the training-node subgraph and testing on the original graph.

Across 10 random Cora training splits with fixed validation and test sets, SGC reports 80.62% test accuracy, close to GCN's 80.53%; GAT is highest among the listed models at 82.29%.

SGC reports 2.70 seconds training time on Reddit, versus 78.54-486.53 seconds for the listed GraphSAGE variants and 270.45 seconds for FastGCN; Figure 3 reports 29×, 180×, 32×, and 100× relative times respectively.

For unseen ImageNet categories only, MLP-SGCZ slightly improves over the authors' GCNZ at both 2-hop and 3-hop accuracy with 4.3M versus 9.5M parameters, but remains below EXEM 1-nns and ADGPM, which use stronger visual features.

Across Cora, Citeseer, and Pubmed, SGC has the lowest reported mean training time among the models in Table 8; Figure 3 labels Pubmed competitors as 6× to 909× SGC's relative training time while SGC retains competitive accuracy.

Replacing DCGCN's GCN with SGC yields 71.0% on NCI1 and 76.2% on COLLAB, described as on par with the GCN counterpart but far behind GIN; the paper gives no GCN or GIN numbers here.

On Cora, SGC and GCN have similar validation accuracy across 1-80 training examples per class; SGC is numerically higher through 40 examples and GCN is higher at 80. The prose specifically says SGC outperforms GCN with fewer than 5 samples.

In the K=2 through K=10 propagation ablation, augmented propagation matrices with self-loops attain higher validation accuracy and more stable performance across depths than their unaugmented counterparts.

K=2 SGC matches or slightly exceeds TextGCN accuracy on 20NG, R8, R52, and Ohsumed while training much faster; on MR it is 0.4 percentage points lower, though still faster. The largest reported speedup is 83.6×.

The authors report that reducing FastGCN's early-stopping window to 10 epochs improved its performance and allowed shorter training time; they caution that FastGCN's originally published CPU benchmark is difficult to compare numerically with their GPU benchmark.

Under Rahimi et al.'s framework, SGC reports higher within-161-mile accuracy and lower training time than GCN+H on GEOTEXT, TWITTER-US, and TWITTER-WORLD, saving more than 30 hours on TWITTER-WORLD.

On QM8, SGC reports 0.03 MAE, substantially worse than the 0.01 MAE reported for both AdaLNet and LNet.

On TACRED, replacing the GCN in C-GCN with K=2 SGC yields C-SGC at 67.0% reported test accuracy versus 66.4% for both the cited and author-run C-GCN, which the paper describes as a new state of the art.

Across 10 random Pubmed training splits with fixed validation and test sets, SGC reports 77.02% test accuracy, close to GCN's 77.09%; GAT is highest among the listed models at 78.79%.

The authors empirically report that normalizing Reddit features to 'zero mean and univariate' is crucial for SGC.

The authors report that GIN performs slightly worse because of overfitting and that LNet and AdaLNet are unstable on citation networks; daggered LNet/AdaLNet statistics exclude low-accuracy outliers.

First-order Chebyshev propagation validation accuracy tends to deteriorate as K increases; the authors attribute this to large low-frequency filter coefficients degrading semi-supervised SGC performance.

For the stated class of undirected weighted simple graphs, adding positive self-loops leaves the smallest normalized-Laplacian eigenvalue at zero and strictly decreases the largest eigenvalue; on Cora/citation-network examples the paper reports an approximate shrinkage from 2 to 1.5, enabling augmented SGC propagation to act as a low-pass-type filter.

Odd propagation depths cause a significant validation-accuracy drop for unaugmented normalized-adjacency and random-walk propagation, whereas adding self-loops minimizes the negative high-frequency coefficients and prevents this degradation.

When evaluation includes unseen and seen ImageNet categories, MLP-SGCZ reports 10.5% at 2 hops and 2.5% at 3 hops, above the authors' GCNZ and the listed GCNZ/ADGPM baselines.

The authors report that an MLP feature extractor is necessary to map pretrained GloVe category vectors into the ResNet-50 visual-feature space before SGC.

On Reddit, SGC reports 94.9% test micro-F1, above supervised SAGE-GCN and FastGCN by more than one point and above both trained and randomly initialized DGI, while below GaAN and two supervised GraphSAGE variants.

Across 10 random Citeseer training splits with fixed validation and test sets, SGC reports 71.40% test accuracy; GAT reports 72.6%, while SGC's reported variation is the largest among the listed non-daggered models.

On the Citeseer public split, SGC reports 71.9% mean test accuracy, about one percentage point above the authors' GCN result; Table 2 reports all listed baselines.

On the Cora public split, SGC reports 81.0% mean test accuracy; Table 2 also reports all listed literature and author-run baselines.

The paper specifies public transductive citation splits, Adam at learning rate 0.2, 60 validation-set hyperopt trials for weight decay, and 100 training epochs; the same paper says epoch count was tuned, while the fixed repository README explicitly requests 150 epochs for Citeseer, so the historical Citeseer Table 2 epoch condition is source-conflicted.

On the Pubmed public split, SGC reports 78.9% mean test accuracy, close to the authors' 79.0% GCN result; Table 2 reports all listed baselines.