Research map / Graph learning / Graph self-supervised learning
Bootstrapping Heterogeneous Graph Representation Learning via Large Language Models: A Generalized Approach
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 GHGRL, which uses Llama 3 to infer node format/content types and descriptions from raw node text, then runs a type-adaptive GNN (PAGNN) for heterogeneous graph representation learning without type information.
Problem
Heterogeneous graph representation learning typically requires known node/edge types and unified node feature formats, which are often unavailable in real-world data with diverse, unstructured attributes.
Approach
An LLM automatically generates format and content type sets, classifies each node with confidence scores and textual descriptions/reasons, and embeds them; a Parameter Adaptive GNN (PAGNN) applies type-specific transformations weighted by LLM confidence, so no type information or preprocessing is needed.
Key results
- On IMDB with 10% training labels, GHGRL reaches 69.73 Macro-F1 versus 61.35 for the best HGNN baseline (PSHGCN) and 61.28 for HAN.
- On IMDB-RIR (r=20%), GHGRL achieves 75.15 Macro-F1, ahead of TAPE (48.59) and GOFA (33.18).
- On Wiki-CS, GHGRL achieves 80.69 Macro-F1 versus 78.65 for GOFA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GHGRL instead of the strongest reported HGNN baseline | GHGRL · IMDB (10% and 40% Training) | Macro-F1 and Micro-F1 | Helped |
| Use GHGRL instead of the strongest reported HGNN baseline | GHGRL · DBLP (10% and 40% Training) | Macro-F1 and Micro-F1 | Hurt |
| Use GHGRL instead of the strongest reported HGNN baseline | GHGRL · ACM (10% and 40% Training) | Macro-F1 and Micro-F1 | Helped |
| Add confidence scores to the type-conditioned processing (ablation) | GHGRL · IMDB (20%, 40%, and 80% Training) | Macro-F1 | Helped |
| Add node-type reasoning descriptions to the LLM-derived features (ablation) | GHGRL · IMDB (20%, 40%, and 80% Training) | Macro-F1 | Helped |
| Use estimated node types rather than ground-truth types (ablation) | GHGRL · IMDB, DBLP, and ACM | Macro-F1 and Micro-F1 | No clear effect |
Surprising findings
- On IMDB, GHGRL with ground-truth types performs only marginally better than GHGRL without type information (70.42 vs 70.34 Macro-F1), and the LLM's type estimates do not fully match true types.
- On DBLP and ACM at 10% training, GHGRL is below SeHGNN/PSHGCN on several metrics, so the claim of best performance is not uniform.
Limitations
- LLM type estimates are imperfect and the method does not accurately recover true node types.
- Theoretical analysis is a simplified linear model and the over-smoothing argument relies on a simplified GHGRL variant.
- Evaluation relies on a single LLM backbone (Llama 3) and does not study other LLMs.
Experimental setup
Datasets: IMDB DBLP ACM Wiki-CS IMDB-RIR DBLP-RID
Benchmarks: IMDB DBLP ACM IMDB-RIR DBLP-RID Wiki-CS
Metrics: Macro-F1 Micro-F1
Baselines: GCN GAT HAN MAGNN SeHGNN PSHGCN TAPE OFA GOFA HAN-w MAGNN-w SHEGNN-w
Builds on: OFA TAPE HAN MAGNN SeHGNN
Closest papers by meaning
- An Efficient Subgraph-Inferring Framework for Large-Scale Heterogeneous Graphs (AAAI 2024 · similarity 0.90)
- Contrastive Auxiliary Learning with Structure Transformation for Heterogeneous Graphs (AAAI 2025 · similarity 0.90)
- Heterogeneous Graph Neural Network on Semantic Tree (AAAI 2025 · similarity 0.89)
- HeterSEED: Semantics–Structure Decoupling for Heterogeneous Graph Learning under Heterophily (NeurIPS 2026 · similarity 0.89)
- HGOT: Self-supervised Heterogeneous Graph Neural Network with Optimal Transport (ICML 2025 · similarity 0.89)
- Label-free Node Classification on Graphs with Large Language Models (LLMs) (ICLR 2024 · similarity 0.89)
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.