LayoutAgent: LLM-Guided Multi-Objective Optimization for Graph Layout Generation
Authors
Yilei He (University of Electronic Science and Technology of China), Jiansu Pu (University of Electronic Science and Technology of China), Chengjin Shi (University of Electronic Science and Technology of China), Miao Zheng (University of Electronic Science and Technology of China), Jinyue Huang (University of Electronic Science and Technology of China)
Presentation
- Session
- I feel tangled in a net
- Time
- Wednesday, Nov 11, 08:36 – 08:48 (US/Eastern) · session 08:00 – 09:30
- Location
- Hall America center
Keywords
Graph Layout, Graph Neural Networks, Multi-Objective Optimization, Adaptive Scheduling, Large Language Models
Abstract
Graph layout is a fundamental problem in graph visualization, where node placement directly affects the readability and interpretability of graph structures. Recent graph neural network (GNN)-based methods have shown promising performance in automatic graph layout generation, but they typically coordinate multiple aesthetic objectives through fixed optimization schemes or weighted loss combinations, making training sensitive to manually specified weights and costly under simultaneous multi-objective optimization. To address these challenges, we propose LayoutAgent, an agent-driven graph layout optimization framework that reformulates multi-objective layout learning as adaptive temporal scheduling. LayoutAgent combines Multi-Objective Aesthetic Graph Embedding (MAGE), enhanced by a Spatial Weighting Adaptation Network (SWAN), with a Phase-wise Multi-Objective Optimization (PMO) mechanism that activates one objective in each training phase. LayoutAgent-Online generates state-dependent schedules to determine the frequency and execution order of aesthetic objectives, while LayoutAgent-Offline distills the online scheduling behavior through LoRA fine-tuning to reduce repeated online LLM queries while preserving comparable scheduling behavior and layout quality. Experiments on the Rome benchmark and representative SuiteSparse graphs show that LayoutAgent achieves competitive layout quality across multiple aesthetic metrics, improves over representative fixed and stochastic scheduling policies, and reduces per-epoch computational cost compared with joint multi-objective optimization baselines.
For Practitioners
This work targets practitioners who use graph visualization for analyzing complex relational data, including data scientists, network analysts, and domain experts in areas such as bioinformatics and knowledge graphs. By automatically balancing multiple layout objectives, LayoutAgent reduces manual tuning efforts and helps practitioners generate clearer and more informative graph representations for data exploration.