GP2F: Cross-Domain Graph Prompting with Adaptive Fusion Framework

This work, proposed by Tianjin University team, addresses the domain gap issue in pre-trained GNN transfer via cross-domain graph prompt learning and adaptive fusion. It has been accepted by ICML 2026.

Graph Neural Networks (GNNs) excel at graph data but face labeled data scarcity. Pre-trained GNNs offer a solution, yet domain gaps between pre-training and downstream tasks limit transfer. Existing methods (fine-tuning, prompting) struggle with cross-domain scenarios like molecular graphs, social networks, and knowledge graphs due to structural and task differences.

GP2F's core idea is to use learnable graph prompts for domain adaptation and adaptive fusion of multiple pre-trained models.

1. Cross-domain Graph Prompt Learning: Adds learnable prompt nodes/edges to input graphs, optimizing to bridge domain gaps without task labels. It's parameter-efficient, preserving pre-trained knowledge.

2. Adaptive Fusion: Dynamically adjusts weights of pre-trained models based on input data characteristics, selecting optimal knowledge combinations for each sample.

Built on PyTorch and PyTorch Geometric, GP2F includes:

• Prompt Encoder: Combines raw graphs with learnable prompts.
• Pre-trained Model Pool: Supports GCN, GAT, GraphSAGE, etc.
• Fusion Network: Adaptive weighted fusion of model outputs.
• Task Head: Task-specific prediction module.

Training uses two stages: first optimize prompt parameters, then jointly optimize prompts and fusion weights to avoid overfitting.

Evaluated on cross-domain benchmarks:

1. Outperforms traditional fine-tuning in cross-domain scenarios (e.g., pre-trained on molecular graphs, downstream social network tasks).
2. Adaptive fusion beats single-model prompting and simple average.
3. Parameter efficiency shines in few-shot learning, achieving near full-supervision performance with limited labels.

Accepted by ICML 2026, reflecting academic recognition.

Applicable to molecular computing, social networks, recommendation systems, drug discovery, and knowledge graph reasoning.

Insights:

• Input-level adaptation (prompt-first) balances efficiency and transfer effect better than forced fine-tuning.
• Adaptive fusion of pre-trained models is more effective than simple ensemble methods.

GP2F represents a key advance in GNN transfer learning. Its cross-domain prompt learning and adaptive fusion address domain gaps effectively. As graph learning applications expand, such flexible domain-adaptive methods will grow in importance.