Research map / Graph learning / Graph self-supervised learning
TGLsta: Low-resource Textual Graph Learning with Semantic and Topological Awareness via LLMs
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 TGLsta, a graph-grounded multi-view contrastive pre-training plus graph prompt tuning method for low-resource (zero- and few-shot) textual graph node classification.
Problem
Low-resource node classification on textual graphs, where labeled data is scarce and existing methods either use shallow text features or fail to fuse semantic and topological information.
Approach
Jointly pre-trains a text encoder, a topology encoder (on random-walk-with-restart subgraphs, using Laplacian eigenvectors), and a GCN graph encoder with text-summary, node-topology, and cross-modal contrastive losses, then adapts the frozen model with learnable prompt vectors for zero- and few-shot classification.
Key results
- Five-shot accuracy of 82.26% on Cora, 88.23% on Industrial, 83.10% on Art and 87.16% on M.I., outperforming the best baseline G2P2 by roughly 2-5 points.
- Zero-shot accuracy of 65.48% on Cora, 79.93% on Industrial, 78.32% on Art and 79.68% on M.I., above G2P2 on all four datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use TGLsta instead of G2p2 | TGLsta · Cora five-shot classification | Accuracy: 80.14 -> 82.26 | Helped |
| use TGLsta instead of G2p2 | TGLsta · Industrial five-shot classification | Accuracy: 82.78 -> 88.23 | Helped |
| use TGLsta instead of G2p2 | TGLsta · Art five-shot classification | Accuracy: 81.08 -> 83.10 | Helped |
| use TGLsta instead of G2p2 | TGLsta · M.I. five-shot classification | Accuracy: 82.71 -> 87.16 | Helped |
| add text-summary contrastive loss Lte−su (ablation) | TGLsta · Cora, Industrial, Art, and M.I. | Accuracy: 80.33 -> 82.26; 86.43 -> 87.12; 82.02 -> 83.10; 86.82 -> 87.16 | Helped |
| add node-structure and summary-structure contrastive losses (ablation) | TGLsta · Cora, Industrial, Art, and M.I. | Accuracy: 80.26 -> 82.26; 86.21 -> 87.12; 81.68 -> 83.10; 86.53 -> 87.16 | Helped |
| add text-summary, node-structure, and summary-structure losses (ablation) | TGLsta · Cora, Industrial, Art, and M.I. | Accuracy: 80.29 -> 82.26; 86.13 -> 87.12; 81.36 -> 83.10; 86.16 -> 87.16 | Helped |
| use trainable prompt vectors instead of only label text (ablation) | TGLsta · Cora, Industrial, Art, and M.I. | Accuracy: 79.57 -> 82.26; 86.60 -> 87.12; 80.91 -> 83.10; 85.32 -> 87.16 | Helped |
| use context-based prompt initialization instead of random initialization (ablation) | TGLsta · Cora, Industrial, Art, and M.I. | Accuracy: 80.37 -> 82.26; 86.96 -> 87.12; 82.97 -> 83.10; 87.06 -> 87.16 | Helped |
| use TGLsta instead of G2p2 | TGLsta · Cora zero-shot classification | Accuracy: 65.10 -> 65.48 | Helped |
| use TGLsta instead of G2p2 | TGLsta · Industrial zero-shot classification | Accuracy: 76.59 -> 79.93 | Helped |
| use TGLsta instead of G2p2 | TGLsta · Art zero-shot classification | Accuracy: 76.94 -> 78.32 | Helped |
Limitations
- Evaluated only on node classification over four text-attributed graphs.
- Relies on a CLIP-sized transformer text encoder (63M parameters) rather than a large LLM for the trainable text branch, and the text embedding relies on LLM features from LLaMA2 or GPT.
Experimental setup
Datasets: Cora Arxiv Industrial Art Music Instruments
Benchmarks: Cora Art Industrial Music Instruments
Metrics: accuracy macro-F1
Baselines: GCN Sagesup TextGCN GPT-GNN DGI SAGEself BERT RoBERTa P-Tuning v2 G2P2
Builds on: G2P2 CLIP GCN P-Tuning v2
Closest papers by meaning
- Leveraging Large Language Models for Node Generation in Few-Shot Learning on Text-Attributed Graphs (AAAI 2025 · similarity 0.90)
- Label-free Node Classification on Graphs with Large Language Models (LLMs) (ICLR 2024 · similarity 0.89)
- Harnessing Explanations: LLM-to-LM Interpreter for Enhanced Text-Attributed Graph Representation Learning (ICLR 2024 · similarity 0.88)
- TEG-DB: A Comprehensive Dataset and Benchmark of Textual-Edge Graphs (NeurIPS 2024 · similarity 0.88)
- SSTAG: Structure-Aware Self-Supervised Learning Method for Text-Attributed Graphs (NeurIPS 2025 · similarity 0.88)
- Unifying Text Semantics and Graph Structures for Temporal Text-attributed Graphs with Large Language Models (NeurIPS 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.