Research map / Graph learning / Graph self-supervised learning
ERAlign: Energy-based Representation Alignment of GNNs and LLMs on Text-attributed Graphs
Xianlin Zeng, Fan Xia, Xiangyu Chen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| align at medium interval {4,12,20,28} instead of output only (ablation) | ERAlignGNN · Cora | accuracy (%): 87.14 -> 90.75 | Improved |
| use medium instead of sparse alignment interval (ablation) | ERAlignGNN · PubMed | accuracy (%): 87.88 -> 92.17 | Improved |
| use dense instead of medium alignment interval (ablation) | ERAlignGNN · Cora, PubMed, Photo | accuracy (%): Cora 90.75 -> 90.21; PubMed 92.17 -> 92.47 | Mixed |
| use Cramer distance with EBM via ED instead of cosine with InfoNCE (ablation) | ERAlignGNN · PubMed | accuracy (%): 89.19 -> 92.17 | Improved |
| use Cramer distance instead of Euclidean distance (EBM via ED) (ablation) | ERAlignGNN · Cora | accuracy (%): 87.32 -> 90.75 | Improved |
| use Cramer distance instead of Wasserstein with Sinkhorn OT (ablation) | ERAlignGNN · Photo | accuracy (%): 88.40 -> 89.38 | Improved |
| train EBM with ED instead of CD (ablation) | ERAlignGNN (Cramer) · Cora | accuracy (%): 90.50 -> 90.75 | Improved |
| train EBM with ED instead of SM (ablation) | ERAlignGNN (Cramer) · Cora | accuracy (%): 88.35 -> 90.75 | Improved |
| add w-stabilization (w=0.1) to ED (ablation) | ERAlign · Cora | training loss convergence | Improved |
| increase fusion strength alpha from 0.05 to 0.5 (ablation) | ERAlignGNN · Cora | accuracy (%): 87.52 -> 90.75 | Improved |
| increase fusion strength alpha from 0.5 to 0.95 (ablation) | ERAlignGNN · Cora | accuracy (%): 90.75 -> 89.27 | Hurt |
| increase perturbed samples M beyond 8 (ablation) | ERAlign · Cora, CiteSeer | accuracy | No clear effect |
Closest papers by meaning
- A Unified Graph Language Model for Multi-Domain Multi-Task Graph Alignment Instruction Tuning (NeurIPS 2026 · similarity 0.91)
- Generalization Principles for Inference over Text-Attributed Graphs with Large Language Models (ICML 2025 · similarity 0.89)
- Multi-View Empowered Structural Graph Wordification for Language Models (AAAI 2025 · similarity 0.89)
- Unifying Text Semantics and Graph Structures for Temporal Text-attributed Graphs with Large Language Models (NeurIPS 2025 · similarity 0.88)
- SSTAG: Structure-Aware Self-Supervised Learning Method for Text-Attributed Graphs (NeurIPS 2025 · similarity 0.88)
- Harnessing Explanations: LLM-to-LM Interpreter for Enhanced Text-Attributed Graph Representation Learning (ICLR 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
