Research map / Graph learning / Graph self-supervised learning
Label-free Node Classification on Graphs with Large Language Models (LLMs)
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 LLM-GNN, a label-free node classification pipeline where GPT-3.5 annotates a small difficulty-aware selected node set with confidence scores, followed by optional post-filtering, and a GCN is trained on those noisy annotations.
Problem
Node classification on text-attributed graphs normally requires abundant ground-truth labels, which are costly and error-prone to obtain; LLM zero-shot prediction on graphs is costly and weaker than trained GNNs.
Approach
Use LLMs as annotators on a small actively selected node set, selecting nodes with a difficulty-aware heuristic (C-Density, based on distance to k-means cluster centers in SBERT feature space) and filtering annotations by LLM-reported confidence and label-entropy change, then train a GCN on these annotations.
Key results
- LLM-GNN reaches 74.91% accuracy on Ogbn-products at about $0.74 cost, versus 75.33% for LLMs-as-Predictors at about $1572.
- On Ogbn-arxiv, LLM-GNN achieves 66.32% accuracy at $0.63, below LLMs-as-Predictors at 73.33% with $79 cost.
- Best variant DA-AGE-W reaches 74.96% on Cora and PS-FeatProp-W reaches 76.23% on Cora under the 20-per-class budget.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LLM-GNN rather than TAG-Z | Ogbn-arxiv label-free node classification · Ogbn-arxiv | accuracy: 37.08 -> 66.32 | Helped |
| Use LLM-GNN rather than LLMs-as-Predictors | Ogbn-products label-free node classification · Ogbn-products | cost (dollars): 1572 -> 0.74 | Helped |
| Replace vanilla zero-shot annotation with one-shot vanilla prompting (ablation) | LLM annotation · Ogbn-products | accuracy: 75.33 ± 4.99 -> 78.67 ± 4.50 | Helped |
| Replace vanilla zero-shot annotation with one-shot vanilla prompting (ablation) | LLM annotation · Ogbn-products | cost (relative to zero-shot): 1 -> 1.8 | Hurt |
| Add post-filtering to FeatProp selection (ablation) | GCN node classifier · Ogbn-products | accuracy: 74.04 ± 0.15 -> 74.91 ± 0.17 | Helped |
| Add weighted cross-entropy to FeatProp selection (ablation) | GCN node classifier · Ogbn-arxiv | accuracy: 66.06 ± 0.07 -> 66.32 ± 0.15 | Helped |
| Combine graph active selection with C-Density (ablation) | GCN with AGE selection · Cora | accuracy: 69.15 ± 0.38 -> 74.38 ± 0.24 | Helped |
| Add weighted cross-entropy to DA-AGE (ablation) | GCN with AGE and C-Density selection · Cora | accuracy: 74.38 ± 0.24 -> 74.96 ± 0.22 | Helped |
| Add post-filtering to DA-AGE (ablation) | GCN with AGE and C-Density selection · Cora | accuracy: 74.38 ± 0.24 -> 71.53 ± 0.19 | Hurt |
| Combine graph active selection with C-Density (ablation) | GCN with RIM selection · Cora | accuracy: 69.86 ± 0.38 -> 73.99 ± 0.44 | Helped |
| Add weighted cross-entropy to DA-RIM (ablation) | GCN with RIM and C-Density selection · Cora | accuracy: 73.99 ± 0.44 -> 74.73 ± 0.41 | Helped |
| Add post-filtering to DA-RIM (ablation) | GCN with RIM and C-Density selection · Cora | accuracy: 73.99 ± 0.44 -> 72.34 ± 0.19 | Hurt |
Surprising findings
- Selecting nodes purely by C-Density (annotation quality) yields poor GCN performance, e.g. 8.29% on Ogbn-products, because of severe label imbalance (all Pubmed annotations fell into one class).
- LLM annotation noise is structured and asymmetric (class-dependent flips), and training dynamics differ from synthetic uniform noise, showing less overfitting.
Limitations
- Relies on proprietary gpt-3.5-turbo API, and only one prompt strategy (zero-shot hybrid) was used in main experiments due to cost.
- Hyperparameters (alpha, beta weights) were not tuned and set equal, and C-Density alone causes class imbalance.
- On Ogbn-arxiv LLM-GNN lags substantially behind LLMs-as-Predictors in accuracy.
Experimental setup
Datasets: Cora Citeseer Pubmed WikiCS Ogbn-arxiv Ogbn-products
Benchmarks: Cora Citeseer Pubmed WikiCS Ogbn-arxiv Ogbn-products
Metrics: accuracy cost in dollars
Baselines: Random AGE RIM FeatProp GraphPart SES TAG-Z BART-large-MNLI LLMs-as-Predictors
Builds on: GCN graph active learning (AGE, RIM, FeatProp, GraphPart) Sentence-BERT self-consistency prompting TAG-Z
Closest papers by meaning
- When Do LLMs Help With Node Classification? A Comprehensive Analysis (ICML 2025 · similarity 0.92)
- Leveraging Large Language Models for Node Generation in Few-Shot Learning on Text-Attributed Graphs (AAAI 2025 · similarity 0.91)
- Harnessing Explanations: LLM-to-LM Interpreter for Enhanced Text-Attributed Graph Representation Learning (ICLR 2024 · similarity 0.91)
- Preference-driven Knowledge Distillation for Few-shot Node Classification (NeurIPS 2025 · similarity 0.90)
- LoReC: Rethinking Large Language Models for Graph Data Analysis (NeurIPS 2026 · similarity 0.90)
- LLaGA: Large Language and Graph Assistant (ICML 2024 · similarity 0.89)
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.