Research map / Graph learning / GNN architectures

Normalize Then Propagate: Efficient Homophilous Regularization for Few-Shot Semi-Supervised Node Classification

AAAI 2025 · Graph learning › GNN architectures · approach: Graph structure learning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes NormProp, an end-to-end decoupled GNN that L2-normalizes node features onto a hypersphere, propagates them, classifies by cosine similarity to fixed prototypes, and adds a homophilous regularization on high-confidence unlabeled nodes based on an upper bound of the propagated norm.

Problem

Semi-supervised node classification with very few labeled nodes per class, where GNNs overfit and suffer label position bias because many nodes lie outside the loss receptive field.

Approach

Decouple direction and Euclidean norm of propagated hyperspherical node representations: direction encodes class information and norm encodes aggregation consistency; derive an upper bound on the norm (random-walk weights) and use the normalized norm as a homophily-based regularizer on confident unlabeled nodes, without pseudo-labels or multi-stage training.

Key results

• On few-shot benchmarks NormProp reaches 81.07% on Cora-ML 5-shot and 64.87% on Ogbn-arxiv at 2.5% label ratio, the best there in the table.
• On standard Planetoid splits NormProp reports 85.46% on Cora and 74.33% on Citeseer, outperforming listed baselines including Violin and GAM.
• On ogbn-arxiv NormProp's training time (79.45 s) and inference time (0.0239 s) are about 17% and 7% of Meta-PN's respectively.

Measured effects reported in the paper

Surprising findings

• Adding the homophilous regularization gives gains over GCN/SGC of about 3% on Cora and Citeseer even though it uses the same fixed low-pass filter, suggesting the regularizer alone drives the improvement.

Limitations

• Homophilous regularization depends on a confidence threshold tau that trades off quantity and quality of selected unlabeled nodes.
• Large lambda drives trivial, over-smoothed solutions that hurt class separability.
• Theoretical analysis relies on the decoupled GCN form without nonlinearities in propagation.

Experimental setup

Datasets: Cora Citeseer Pubmed Cora-ML MS-CS Ogbn-arxiv

Benchmarks: Cora Citeseer Pubmed Cora-ML MS-CS Ogbn-arxiv

Metrics: test accuracy training time inference time

Baselines: GCN SGC APPNP DAGNN C&S GPR-GNN LP GLP IGCN M3S Meta-PN GRAND

Builds on: APPNP SGC GCN Meta-PN

Closest papers by meaning

• Regularizing Graph Neural Networks via Consistency-Diversity Graph Augmentations (AAAI 2022 · similarity 0.90)
• Graph Random Neural Networks for Semi-Supervised Learning on Graphs (NeurIPS 2020 · similarity 0.90)
• Meta Propagation Networks for Graph Few-shot Semi-supervised Learning (AAAI 2022 · similarity 0.89)
• Beyond Homophily in Graph Neural Networks: Current Limitations and Effective Designs (NeurIPS 2020 · similarity 0.89)
• Unsupervised Neighborhood Propagation Kernel Layers for Semi-supervised Node Classification (AAAI 2024 · similarity 0.89)
• Fast Graph Sharpness-Aware Minimization for Enhancing and Accelerating Few-Shot Node Classification (NeurIPS 2024 · similarity 0.88)