Where LLM Annotators Fail: Label-Free Learning on Graphs with LLMs

2026-05-26

AI Summary

This work addresses the challenge of noisy pseudo-labels generated by large language models (LLMs) for graph nodes, a problem exacerbated by existing methods’ neglect of the dual dependency of noise on both class identity and local feature-space regions. To this end, we propose CANE, an unsupervised graph learning framework that, for the first time, models cluster-aware label reliability without access to ground-truth labels. By dynamically selecting or refining pseudo-labels based on their contextual consistency within learned clusters, CANE effectively enhances node classification performance. The approach seamlessly integrates graph neural networks, LLM-generated pseudo-labels, unsupervised clustering, and noise-robust learning principles. Extensive experiments demonstrate that CANE significantly outperforms state-of-the-art label-free methods across multiple graph benchmarks and GNN backbones, with the largest gains observed on datasets exhibiting pronounced cluster-conditional noise patterns.

Abstract

Node classification on graphs often requires labeled nodes, yet obtaining labels at graph scale is expensive. When node attributes contain semantic content, such as paper abstracts, web pages, or product descriptions, large language models (LLMs) can provide low-cost supervision by annotating a small subset of nodes. However, these LLM-generated labels are noisy, and existing label-free graph learning methods usually treat this noise as either global or class-conditional. We find that LLM annotation errors are not only class-dependent but also region-dependent: within the same class, reliability can vary sharply across feature-space clusters. In light of this, we propose Cluster-Aware Noise Estimation (CANE), a label-free learning framework that estimates cluster-conditional LLM reliability without ground truth labels, and uses this estimate to decide which pseudo-labels to trust, and which labels to correct. Across various graph benchmarks and GNN backbones, CANE improves over the strongest label-free baselines, with the largest gains on datasets exhibiting stronger cluster-conditional noise.

Problem

• LLM annotation noise
• node classification
• label-free learning
• graph neural networks
• cluster-conditional noise

Innovation

• LLM annotation noise
• cluster-conditional reliability
• label-free graph learning
• pseudo-label correction
• graph neural networks