AI Paper Atlas

Research map / Graph learning / GNN architectures

Towards Understanding and Reducing Graph Structural Noise for GNNs

Mingze Dong, Yuval Kluger

ICML 2023 · Graph learning › GNN architectures · approach: Structural GNN variants

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
GPS graph rewiring (propensity-score based edge reweighting, ensembled with original graph)GCN · WebKB (Cornell, Texas, Wisconsin) and Actor, Chameleon*node classification accuracy: Cornell 43.5 to 74.6; Texas 66.2 to 80.0; Wisconsin 52.9 to 77.3Improved
GPS graph rewiringGCN · Homophilic datasets (Cora, Citeseer, Pubmed)node classification accuracy: Cora 81.7 to 80.5; Citeseer 71.8 to 71.5; Pubmed 78.9 to 77.7Hurt
k-NN graph rewiring on data featuresGCN · Cornellnode classification accuracy: 43.5 to 68.6Improved
k-NN graph rewiring on data featuresGCN · Coranode classification accuracy: 81.7 to 48.0Hurt
GPS-rewired graph ESNR relationship to gainsGPS rewired graphs · Node classification benchmarkscorrelation of ESNR with accuracy gainMixed
Edge dropout ratio increase (ESNR tracked) (ablation)Cora, Citeseer, Pubmed with GCN · Planetoid edge dropoutESNR and GCN accuracyImproved
ESNR as metric (vs edge homophily) for GCN minus MLP accuracy difference (ablation)14 datasets · Cross-dataset GCN vs MLPPearson r: ESNR vs GCN r=0.85 vs EH vs GCN r=0.58Improved
Vanilla entry-wise SNR instead of ESNR (ablation)Simulated CSBM · CSBM edge ratio sweepconsistency with GCN trendHurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy