Efficient Text-Attributed Graph Learning through Selective Annotation and Graph Alignment

Huanyi Xie, Lu Yu, Tianhao Huang, Longfei Li, Meng Li, JUN ZHOU, Lijie Hu, Di Wang

Abstract:

In the realm of Text-attributed Graphs (TAGs), traditional graph neural networks (GNNs) often fall short due to the complex textual information associated with each node. Recent methods have improved node representations by leveraging large language models (LLMs)
to enhance node text features, but these approaches typically require extensive annotations or fine-tuning across all nodes, which is both time-consuming and costly. To overcome these challenges, we introduce GAGA, an efficient framework for TAG representation learning.
GAGA reduces annotation time and cost by focusing on annotating only representative nodes and edges. It constructs an annotation graph that captures the topological relationships among these annotations. Furthermore, GAGA employs a two-level alignment module to effectively integrate the annotation graph with the TAG, aligning their underlying structures. Experiments show that GAGA achieves classification accuracy on par with or surpassing state-of-the-art methods while requiring only 1% of the data to be annotated, demonstrating its high efficiency.

Submission Type: Regular submission (no more than 12 pages of main content)

Changes Since Last Submission:

Summary of Changes in the Revised Manuscript
We thank all reviewers for their constructive feedback. Below we summarize the changes made in the revised version (highlighted in blue in the PDF).
1. New Section: Limitations and Future Work (Section 6)
We added a dedicated section discussing the limitations of GAGA and corresponding future directions, covering:

Scope of evaluation: discussion of the transductive setting, extension to high-heterophily graphs, and long-tail class distributions.
Dependence on LLM annotation quality: connection to our noise robustness analysis, and applicability to privacy-constrained domains using locally deployable open-source LLMs.
Beyond text-attributed graphs: extension to heterogeneous/dynamic TAGs and non-text-attributed graphs via multimodal LLMs.

2. Updated Related Work (Section 2)
We extended the Related Work section to cover recent LLM-enhanced TAG methods from the last two years, specifically UniGLM (Fang et al., 2024) and AuGLM (Xu et al., 2026), and clarified how GAGA differs from them: GAGA only requires annotating a small fraction of nodes and aligns them with the graph through a lightweight encoder, rather than encoding the full node set with a large language model.
3. New Appendix B.1: Comparison with Recent LLM-Enhanced TAG Methods
We added a direct comparison (Table 13) against UniGLM and AuGLM on ogbn-arxiv, PubMed, ogbn-products, and Cora. Notably, GAGA outperforms AuGLM on ogbn-arxiv (76.21 vs. 76.00) while using a text encoder that is roughly 34× smaller (all-MiniLM-L6-v2, 22.7M parameters vs. Flan-T5-Large, 770M parameters).
4. New Experiment: Robustness to Annotation Noise (Appendix B)
We added a controlled noise-injection study on ogbn-arxiv (Table 12), where annotations between selected nodes are randomly shuffled at rates of 0%, 10%, and 20%. The results show that the alignment loss rises consistently (1.272 → 1.354 → 1.569) and accuracy degrades gracefully, confirming that GAGA genuinely learns from the annotation signal rather than overfitting to spurious patterns.
5. New Experiment: Feature Combination Ablation (Appendix B)
We added an ablation study (Table 11) comparing different feature combination strategies in the downstream fine-tuning stage—cross-attention, concatenation, and averaging—against the GCN baseline. The results demonstrate that cross-attention (76.27% test accuracy) substantially outperforms concatenation (71.08%) and averaging (71.98%), justifying our design choice.
6. New Table: GNN Backbone Configurations Across Methods (Appendix A.3.1)
We added Table 7, which lists the GNN backbone, number of layers, and hidden dimension for every baseline. This clarifies that GAGA and the closest LLM-enhanced baseline (TAPE) use the same 4-layer, 128-hidden-dim GCN, so GAGA's improvement over TAPE is not an artifact of GNN depth or width.
7. Clarification of the "1%" Annotation Ratio (Appendix A.3.1)
We explicitly clarified that the "1%" refers to the fraction of nodes that receive LLM-generated annotations for the alignment stage, not the fraction of labeled training data. Train/validation/test splits remain standard and identical across all methods.

Assigned Action Editor: ~Xiaofeng_Cao1

Submission Number: 7359