Research map / Graph learning / Graph self-supervised learning
Multi-View Empowered Structural Graph Wordification for Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Dr.E, an end-to-end dual-residual VQ-VAE that maps GNN node embeddings (multi-hop views) onto LLM token embeddings so a fine-tuned Llama2-7B can classify graph nodes in natural language.
Problem
Integrating graph-structured data with LLMs: existing approaches either serialize graphs as raw text (losing structure) or inject GNN embeddings into LLMs (losing interpretable prompt semantics).
Approach
Quantize GNN node embeddings into sequences of codes drawn from the LLM's own token embeddings using a residual (RQ-VAE-style) codebook with intra-layer and inter-layer residuals, and feed multi-hop views of the central node as natural-language prompts to the LLM.
Key results
- Dr.E reaches 0.9132 accuracy on Cora, 0.9670 on PubMed, and 0.7645 on Arxiv, averaging 0.8815 and beating LLaGA (0.8697 average) and GraphEdit (0.8692).
- Ablation on Llama2-7B + GNN: adding LoRA raises Cora accuracy from 0.8319 to 0.8824; multi-view further raises it to 0.9112.
- Performance peaks at three views; perplexity per layer keeps a roughly 1:2:3 ratio according to the authors.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add LoRA fine-tuning to the LLM (ablation) | Llama2-7B + GNN · Cora, PubMed, Arxiv | accuracy: Cora: 0.8319 -> 0.8824; PubMed: 0.8756 -> 0.9213; Arxiv: 0.6513 -> 0.7235 | Helped |
| add Multi-View Structural Enhancement (ablation) | Dr.E with LoRA · Cora, PubMed, Arxiv | accuracy: Cora: 0.8824 -> 0.9112; PubMed: 0.9213 -> 0.9524; Arxiv: 0.7235 -> 0.7468 | Helped |
| add quantization and map embeddings to a frozen codebook (ablation) | Dr.E with Multi-View · Cora, PubMed, Arxiv | accuracy: Cora: 0.9112 -> 0.8875; PubMed: 0.9524 -> 0.9482; Arxiv: 0.7468 -> 0.7043 | Hurt |
| add Intra-Layer Residual (ablation) | Dr.E with VQ and frozen codebook · Cora, PubMed, Arxiv | accuracy: Cora: 0.8875 -> 0.9036; PubMed: 0.9482 -> 0.9598; Arxiv: 0.7043 -> 0.7354 | Helped |
| add Inter-Layer Residual (ablation) | Dr.E with Intra-Layer Residual · Cora, PubMed, Arxiv | accuracy: Cora: 0.9036 -> 0.9101; PubMed: 0.9598 -> 0.9655; Arxiv: 0.7354 -> 0.7509 | Helped |
| refine the codebook by filtering tokens without meaningful semantics (ablation) | Dr.E with dual residuals · Cora, PubMed, Arxiv | accuracy: Cora: 0.9101 -> 0.9132; PubMed: 0.9655 -> 0.9670; Arxiv: 0.7509 -> 0.7645 | Helped |
| increase the number of structural views (ablation) | Dr.E · node classification | relative performance | Helped |
Surprising findings
- Dr.E is reported to be slightly below LLaGA on OGBN-Arxiv, and quantization/token mapping initially decreases performance (0.9112 to 0.8875 on Cora).
Limitations
- Computational requirements limit scaling to extremely large and complex graphs.
- Improvement over other models is modest, and it trails LLaGA on OGBN-Arxiv.
- Generalization to non-textual graphs is not yet tested on additional datasets.
Experimental setup
Datasets: Cora PubMed OGBN-Arxiv
Benchmarks: Cora PubMed OGBN-Arxiv
Metrics: classification accuracy
Baselines: GCN GraphSAGE GAT RevGAT DRGAT SGC SAGN Graphformer Nodeformer InstructGLM GLEM GraphMAE2
Builds on: VQ-VAE RQ-VAE LQAE VGAE GraphSAGE
Closest papers by meaning
- A Unified Graph Language Model for Multi-Domain Multi-Task Graph Alignment Instruction Tuning (NeurIPS 2026 · similarity 0.91)
- LLaGA: Large Language and Graph Assistant (ICML 2024 · similarity 0.90)
- UniGTE: Unified Graph–Text Encoding for Zero-Shot Generalization across Graph Tasks and Domains (NeurIPS 2025 · similarity 0.89)
- GOFA: A Generative One-For-All Model for Joint Graph Language Modeling (ICLR 2025 · similarity 0.88)
- Generalization Principles for Inference over Text-Attributed Graphs with Large Language Models (ICML 2025 · similarity 0.88)
- GRAPHGPT-O: Synergistic Multimodal Comprehension and Generation on Graphs (CVPR 2025 · 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.