[論文レビュー] MAGPrompt: Message-Adaptive Graph Prompt Tuning for Graph Neural Networks

Long D. Nguyen, Binh P. Nguyen|arXiv (Cornell University)|Feb 5, 2026

Advanced Graph Neural Networks被引用数 0

ひとこと要約

MAGPrompt は事前学習済み GNN バックボーンをフリーズし、隣接メッセージを再重み付けし、集約時に加法的信号を注入する軽量プロンプトを学習して、Few-shot 設定で下流パフォーマンスを向上させつつ、全ファインチューニングと競合可能な状態を維持する。

ABSTRACT

Pre-trained graph neural networks (GNNs) transfer well, but adapting them to downstream tasks remains challenging due to mismatches between pre-training objectives and task requirements. Graph prompt tuning offers a parameter-efficient alternative to fine-tuning, yet most methods only modify inputs or representations and leave message passing unchanged, limiting their ability to adapt neighborhood interactions. We propose message-adaptive graph prompt tuning, which injects learnable prompts into the message passing step to reweight incoming neighbor messages and add task-specific prompt vectors during message aggregation, while keeping the backbone GNN frozen. The approach is compatible with common GNN backbones and pre-training strategies, and applicable across downstream settings. Experiments on diverse node- and graph-level datasets show consistent gains over prior graph prompting methods in few-shot settings, while achieving performance competitive with fine-tuning in full-shot regimes.

研究の動機と目的

• Motivate improving transferability of pre-trained GNNs across downstream tasks without full fine-tuning.
• Develop a parameter-efficient prompting mechanism that directly modulates message passing.
• Ensure compatibility with common GNN backbones and pre-training strategies while keeping the backbone frozen.
• Demonstrate gains over representation-based graph prompting methods, especially in few-shot settings.

提案手法

• Freeze the GNN backbone and learn a small set of prompts per layer.
• Introduce a per-edge message gate aij(l) to reweight neighbor messages during aggregation.
• Add a shared message prompt p(l) injected at the message level during aggregation.
• Optionally extend to MAGPrompt+ with multiple prompt bases and edge-adaptive composition for richer relational patterns.
• Maintain permutation equivariance by preserving graph structure and backbone parameters.
• Provide a prompt-collapse regularization to balance usage of prompt bases in MAGPrompt+.

実験結果

リサーチクエスチョン

• RQ1 Does message-adaptive prompting improve downstream adaptation in few-shot and full-shot settings compared to representation-based prompting?
• RQ2 How do MAGPrompt and MAGPrompt+ components (reweighting and compositional edge prompts) contribute to performance across datasets and pre-training strategies?
• RQ3 How do hyperparameters (prompt basis size, beta, regularization strength, attention dimension) affect results?

主な発見

• MAGPrompt+ achieves the best average performance across 5-shot node classification benchmarks, outperforming prior graph prompting methods.
• MAGPrompt consistently improves over representation-based prompting across datasets and pre-training strategies.
• In graph classification (50-shot), MAGPrompt+ attains the highest average accuracy across all benchmarks and pre-training methods.
• In full-shot MoleculeNet benchmarks, MAGPrompt+ attains state-of-the-art ROC-AUC, often surpassing fine-tuning.
• Ablation studies show message-adaptive reweighting is crucial, and adding message-wise prompts provides additional gains.
• Prompt bases (10-20) offer a balance between expressiveness and efficiency, with regularization mitigating prompt collapse.