AI Paper Atlas

Research map / Graph learning / GNN architectures

Alternately Optimized Graph Neural Networks

Haoyu Han, Xiaorui Liu, Haitao Mao, MohamadAli Torkamani, Feng Shi, Victor Lee, Jiliang Tang

ICML 2023 · Graph learning › GNN architectures · approach: Graph structure learning · 0 citations (OpenAlex)

Read the paper

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
Alternating optimization with lazy propagation (update F only 5 times) instead of per-epoch propagation (ablation)ALT-OPT on ogbn-arxiv · ogbn-arxivaccuracy (%): 72.70 vs 72.76 (Full)No clear effect
Alternating optimization with lazy propagation (update F only 5 times) instead of per-epoch propagation (ablation)ALT-OPT on ogbn-products · ogbn-productsaccuracy (%): 82.64 vs 81.83 (Full)Improved
Reduce F updates to only 1 time during training (ablation)ALT-OPT on ogbn-arxiv · ogbn-arxivaccuracy (%): 70.09 vs 72.70 (ALT-OPT-5)Hurt
Increase F-update frequency from 1 to 5 times (ablation)ALT-OPT on ogbn-products · ogbn-productsaccuracy (%): 80.03 (1x) to 82.64 (5x)Improved
Alternating optimization (ALT-OPT-5) vs APPNP end-to-end trainingogbn-products · ogbn-productstraining time (s): 222 s vs 1,913 sImproved
Alternating optimization with lazy propagation vs APPNPogbn-products · ogbn-productsmemory (GB): 24.49 vs 29.15Improved
Feature diffusion (LP-enhanced features) before MLP pretraining removed (ablation)ALT-OPT on Cora and CiteSeer · Cora, CiteSeer low label ratesaccuracy (%): gap large at low label rates, small at higher ratesImproved
Remove pseudo labels (train MLP on labeled data only) (ablation)ALT-OPT on Cora and CiteSeer · Cora, CiteSeeraccuracy (%): large impact especially at low label rate (figure only)Improved
Replace MSE loss with cross-entropy (ALT-OPT-N with softmax normalization and reweighting) (ablation)ALT-OPT · PubMed, CS, ogbn-arxivaccuracy (%): 78.66 vs 79.16 (PubMed); 92.69 vs 92.77 (CS); 72.58 vs 72.60 (arxiv)No clear effect
Use Laplacian filter L^2 (ALT-OPT-H) instead of L on heterophily graphsALT-OPT on Chameleon · Chameleonaccuracy (%): 70.62 vs LINKX 68.42Improved
Use Laplacian filter L^2 (ALT-OPT-H) instead of L on heterophily graphsALT-OPT on Squirrel · Squirrelaccuracy (%): 61.56 vs LINKX 61.81No clear effect
Alternating optimization with pseudo-labels vs end-to-end GNNs at low label ratesALT-OPT on Cora, CiteSeer, Pubmed · label rate 5 (transductive)accuracy (%): Cora 76.78 vs APPNP 75.86; CiteSeer 67.48 vs 63.92; Pubmed 73.51 vs 72.68Improved

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