Graph prompt learning has emerged as a promising "pretraining–adaptation" paradigm for training graph learning models. However, existing methods for prompt construction typically rely on limited supervision to generate class prototypes, which may compromise the semantic quality of the prompts under few-shot conditions. Prototype-based prompt methods therefore often suffer from biased prototype representations in small-sample scenarios, leading to degraded model performance. To address this issue, we propose a prototype-enhanced graph prompt learning framework. During the pretraining stage, global structural semantics are explicitly incorporated to obtain more robust prompt representations. In the fine-tuning stage, a combination of pairwise constraint construction and pseudo-label augmentation is employed to improve prototype estimation, while semantic consistency regularization enhances the model’s robustness to node and structural perturbations. Experimental results demonstrate that the proposed method consistently outperforms existing prompt-based few-shot graph learning approaches on standard node classification benchmarks and exhibits improved robustness.
Paper
Full text
Prototype-Enhanced Graph Prompt Learning for Few-Shot Node Classification
OpenAlex · Advanced Graph Neural Networks · 2026
Abstract
Graph prompt learning has emerged as a promising "pretraining–adaptation" paradigm for training graph learning models. However, existing methods for prompt construction typically rely on limited supervision to generate class prototypes, which may compromise the semantic quality of the prompts under few-shot conditions. Prototype-based prompt methods therefore often suffer from biased prototype representations in small-sample scenarios, leading to degraded model performance. To address this issue, we propose a prototype-enhanced graph prompt learning framework. During the pretraining stage, global structural semantics are explicitly incorporated to obtain more robust prompt representations. In the fine-tuning stage, a combination of pairwise constraint construction and pseudo-label augmentation is employed to improve prototype estimation, while semantic consistency regularization enhances the model’s robustness to node and structural perturbations. Experimental results demonstrate that the proposed method consistently outperforms existing prompt-based few-shot graph learning approaches on standard node classification benchmarks and exhibits improved robustness.