GNN-as-Judge: Unleashing the Power of LLMs for Graph Learning with GNN Feedback

Researchers propose GNN-as-Judge, a framework combining Large Language Models with Graph Neural Networks to improve learning on text-attributed graphs in low-resource settings. The approach uses collaborative pseudo-labeling and weakly-supervised fine-tuning to generate reliable labels while reducing noise, demonstrating significant performance gains when labeled data is scarce.

GNN-as-Judge addresses a fundamental limitation in machine learning: the performance degradation of LLMs when labeled training data becomes severely restricted. This research matters because it bridges two powerful but complementary technologies—LLMs excel at semantic understanding while GNNs capture structural relationships—to solve a practical problem that affects real-world deployments across knowledge graphs, recommendation systems, and knowledge bases where annotation is expensive.

The framework's innovation centers on a collaborative pseudo-labeling strategy that leverages disagreement between LLMs and GNNs as a signal for data quality. Rather than treating pseudo labels as binary truth, the approach identifies unlabeled nodes most influenced by labeled ones and uses pattern agreement between the two models to determine label reliability. This mitigates a core challenge in semi-supervised learning: the propagation of incorrect pseudo labels through training cycles.

For the AI research community, this work demonstrates how structural inductive biases from graph neural networks can enhance LLM fine-tuning, particularly valuable for industries managing complex networked data. The few-shot semi-supervised setting aligns with practical constraints across finance, social networks, and knowledge graphs where obtaining comprehensive labels remains prohibitively expensive. The weakly-supervised fine-tuning algorithm further reduces reliance on large labeled datasets.

The experimental validation across multiple text-attributed graph datasets suggests this approach could accelerate deployment of LLM-based systems in resource-constrained environments. Future developments might explore cross-domain transfer capabilities and scalability to billion-node graphs, expanding applicability across enterprises managing large knowledge structures.

• →GNN-as-Judge combines LLMs and GNNs to improve performance in low-resource graph learning scenarios with limited labeled data.
• →The collaborative pseudo-labeling strategy uses agreement and disagreement patterns between models to generate reliable training labels.
• →Weakly-supervised fine-tuning mitigates label noise while distilling knowledge from informative pseudo labels.
• →The framework significantly outperforms existing methods particularly in few-shot semi-supervised settings.
• →This approach has applications across knowledge graphs, recommendation systems, and networked data domains requiring cost-effective annotation.