Graph Transformers Show Size Transferability, Matching GNNs

Researchers have established a theoretical connection between Graph Transformers (GTs) and Manifold Neural Networks (MNNs), particularly when GTs utilize GNN-based positional encodings. This study demonstrates that GTs inherit transferability guarantees from their positional encodings, meaning models trained on smaller graphs can generalize to larger ones. Experimental results on standard benchmarks confirm that GTs exhibit scalable behavior comparable to GNNs, with practical applications shown in shortest path distance estimation over terrains. AI

IMPACT Demonstrates theoretical underpinnings for GTs' scalability, suggesting efficient training methods for large-scale graph data.

RANK_REASON Academic paper detailing theoretical findings and experimental validation of graph transformer properties. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• Graph Neural Network
• Graph Transformers
• Javier Porras-Valenzuela
• Manifold Neural Networks

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →