EvoPath: Evolutionary meta-path discovery with large language models for complex heterogeneous information networks

2025-01-01

Information Processing & Management

AI Summary

To address the low efficiency, poor generalizability, and weak interpretability of manually designed meta-paths in heterogeneous information networks (HINs), this paper proposes the first automated meta-path discovery framework integrating large language model (LLaMA-3) semantic reasoning with a multi-objective evolutionary algorithm (NSGA-II). Our approach innovatively embeds LLM-based semantic understanding into the evolutionary search process, introduces a differentiable path evaluator for end-to-end optimization, and supports cross-domain adaptation and task-driven evolution. Evaluated on five standard HIN benchmarks, the framework achieves an average 12.7% improvement in link prediction F1-score, enhances path interpretability by 3.2× (measured via human-evaluated interpretability scores), and accelerates meta-path generation by 20× compared to manual design.

Technology Category

• Search and Optimization: Metareasoning and Metaheuristics
• Natural Language Processing: Information Extraction
• Machine Learning: Large Multimodal Models (LMMs)

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics

Problem

• Heterogeneous Information Networks
• Meta-paths Optimization
• Large Language Models

Innovation

• EvoPath
• Enhanced Meta-path Search
• Large Language Models