GNN performance on heterophilic graphs depends on node representations

A new research paper explores the performance of Graph Neural Networks (GNNs) on heterophilic graphs, where connected nodes often have dissimilar labels. The study found that the effectiveness of different GNN architectures is significantly dependent on the type of node representations used. For instance, on the Roman-Empire benchmark, the performance gain from using contextual Transformer embeddings over static fastText vectors varied substantially across architectures, with GAT showing a larger improvement than GCN-sep. This suggests that architectural choices and representation strategies cannot be evaluated in isolation when assessing GNN robustness in heterophilic settings. AI

IMPACT Highlights the critical interplay between GNN architecture and node representation for effective performance on heterophilic graphs.

RANK_REASON The cluster contains a research paper detailing findings on Graph Neural Networks. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.IR (Information Retrieval) →

• Amazon-Ratings
• fastText
• GCN-sep
• graph attention network
• graph convolutional network
• graph neural networks
• H2GCN
• MPNet: Masked and Permuted Pre-training for Language Understanding
• Roman-Empire
• transformer

• paper
• other

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