When LLM Agents Meet Graph Optimization: An Automated Data Quality Improvement Approach

TL;DR

This work tackles the data quality bottleneck in text-attributed graphs (TAGs) by introducing LAGA, a unified, automated framework powered by large language models. LAGA casts TAG quality improvement as a closed-loop, data-centric process implemented via four collaborating agents—Detection, Planning, Action, and Evaluation—that detect multi-modal defects, plan adaptive repairs, apply text/structure/label enhancements, and assess improvements. Empirical results across five datasets, nine degradation scenarios, and multiple backbones show that LAGA consistently achieves state-of-the-art performance and robustness, aided by scalability strategies like edge sampling and subgraph partitioning. The approach demonstrates that holistic TAG quality optimization is crucial for reliable graph analytics and lays groundwork for extending to heterogeneous or more complex graph settings.

Abstract

Text-attributed graphs (TAGs) have become a key form of graph-structured data in modern data management and analytics, combining structural relationships with rich textual semantics for diverse applications. However, the effectiveness of analytical models, particularly graph neural networks (GNNs), is highly sensitive to data quality. Our empirical analysis shows that both conventional and LLM-enhanced GNNs degrade notably under textual, structural, and label imperfections, underscoring TAG quality as a key bottleneck for reliable analytics. Existing studies have explored data-level optimization for TAGs, but most focus on specific degradation types and target a single aspect like structure or label, lacking a systematic and comprehensive perspective on data quality improvement. To address this gap, we propose LAGA (Large Language and Graph Agent), a unified multi-agent framework for comprehensive TAG quality optimization. LAGA formulates graph quality control as a data-centric process, integrating detection, planning, action, and evaluation agents into an automated loop. It holistically enhances textual, structural, and label aspects through coordinated multi-modal optimization. Extensive experiments on 5 datasets and 16 baselines across 9 scenarios demonstrate the effectiveness, robustness and scalability of LAGA, confirming the importance of data-centric quality optimization for reliable TAG analytics.

Figures (7)

• Figure 1: Performance comparison of two GNN backbones (GCN, GAT) and two LLM-GNN backbones (TAPE, ENGINE) across nine TAG quality scenarios on Cora. Each point represents a specific quality degradation case, evaluated on two downstream tasks: node classification accuracy (Node cls acc) and node clustering normalized mutual information (Node clu nmi). For the label imbalance scenario, we report the accuracy and NMI of the minority class under both the original and perturbed graphs.
• Figure 2: The framework of LAGA, including the overall workflow of LAGA and the internal module details of each agent.
• Figure 3: Node clustering NMI comparison across different scenarios with perturbation ratio = 0.4. The "Org" denotes the original graph, while "TS", "TN", "TI", ... represent the 9 types of scenarios defined in Sec. \ref{['section:scene_def']}. "LAGA-" refers to the performance of LAGA in each of these scenarios.
• Figure 4: Performance of LAGA with different backbones across nine scenarios, showing the accuracy improvements over the corresponding backbones (GCN, GAT, GraphSAGE, TAPE and ENGINE).
• Figure 5: The sensitivity of LAGA to different hyperparameters $(k_{edge}, \tau_{edge}, \tau_{lape}, r_{gen})$. "Dataset-Scene" denotes different datasets and experimental scenarios, while "Rel-Acc" refers to the relative accuracy in node classification.