Research map / Graph learning / Graph self-supervised learning
Uncovering Semantic Hierarchies in Text-Attributed Graphs via Variational EM-based LLM–GNN Synergy
Yunhui Liu, Xudong Jin, Qizhuo Xie, Chunhui Zhao, Xiao Luo, Tao Zheng, Bin Chong, Tieke He
Read the paperCheck your idea against 137,217 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 SHiFT, a variational EM framework that couples LLM reasoning with GNN structure encoding to uncover coarse-to-fine semantic hierarchies in text-attributed graphs.
Problem
Representation learning on text-attributed graphs typically uses flat semantic spaces and ignores the coarse-to-fine hierarchical granularity present in real-world data.
Approach
Synergizes LLM reasoning with GNN structural representations inside an Expectation-Maximization (variational EM) loop to infer hierarchical semantic structure over text-attributed graph nodes, rather than a flat semantic space.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add iterative EM refinement of the taxonomy (vs one-pass initial taxonomy) (ablation) | SHiFT · Cora, Citeseer, WikiCS | accuracy: avg drop 1.27 without EM; Cora 84.83 -> 83.27 | Improved |
| add taxonomy partition alignment (TPA) (ablation) | SHiFT · Cora, Citeseer, WikiCS | accuracy: avg drop 1.24 without TPA | Improved |
| add taxonomy skeleton alignment (TSA) (ablation) | SHiFT · Cora, Citeseer, WikiCS | accuracy: avg drop 1.39 without TSA; Citeseer 73.96 -> 72.31 | Improved |
| add taxonomy concept alignment (TCA) (ablation) | SHiFT · Cora, Citeseer, WikiCS | accuracy: avg drop 0.48 without TCA; Cora 84.83 -> 84.14 | Improved |
| induce taxonomy with graph structure instead of PLM embeddings only (ablation) | SHiFT taxonomy · Cora, Citeseer, WikiCS | NPMI: Cora 0.1640 -> 0.1958 | Improved |
| EM refinement of taxonomy (T_init to T_final) (ablation) | SHiFT taxonomy · Cora human evaluation | Hierarchy Validity: 3.65 -> 4.10 | Improved |
| append SHiFT-induced taxonomy path to raw text features (ablation) | MLP classifier · Cora | accuracy: 76.05 -> 81.91 | Improved |
| append SHiFT-induced taxonomy path to raw text features (ablation) | GCN classifier · Cora | accuracy: 81.70 -> 83.41 | Improved |
| append path from graph-agnostic taxonomy to raw text features (ablation) | MLP classifier · Cora | accuracy: 76.05 -> 72.68 | Hurt |
| explicit semantic taxonomy instead of hyperbolic implicit hierarchy | SHiFT vs HGCN/HGCL · node classification avg | accuracy: 77.26 -> 81.40 | Improved |
| Mapping & Diagnosis incremental updates instead of full regeneration | SHiFT E-step · later epochs | LLM invocations: reduces by ~80% | Improved |
Closest papers by meaning
- Generalization Principles for Inference over Text-Attributed Graphs with Large Language Models (ICML 2025 · similarity 0.88)
- Harnessing Explanations: LLM-to-LM Interpreter for Enhanced Text-Attributed Graph Representation Learning (ICLR 2024 · similarity 0.87)
- SSTAG: Structure-Aware Self-Supervised Learning Method for Text-Attributed Graphs (NeurIPS 2025 · similarity 0.86)
- UniGTE: Unified Graph–Text Encoding for Zero-Shot Generalization across Graph Tasks and Domains (NeurIPS 2025 · similarity 0.86)
- Large Language Model Meets Graph Neural Network in Knowledge Distillation (AAAI 2025 · similarity 0.86)
- TGLsta: Low-resource Textual Graph Learning with Semantic and Topological Awareness via LLMs (AAAI 2025 · similarity 0.86)
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.
