More about HKUST
Toward Real-World Graph Intelligence: A Systematic Journey through Data, Architecture, and Applications
PhD Thesis Proposal Defence Title: "Toward Real-World Graph Intelligence: A Systematic Journey through Data, Architecture, and Applications" by Mr. Peiyan ZHANG Abstract: The rapid advancement of graph neural networks (GNNs) has revolutionized machine learning on relational data, enabling state-of-the-art performance in tasks like node classification, link prediction, and graph-level inference. By aggregating information from neighboring nodes and edges, GNNs effectively integrate local topology and attribute information. However, real-world graph intelligence often diverges from controlled benchmarks, presenting challenges such as structural variability, domain shifts, temporal evolution, and multimodal or incomplete attributes. These complexities hinder learning, generalization, and robustness. Additionally, existing GNNs struggle to unify local structural information with global semantics, limiting their ability to model long-range dependencies, avoid over-smoothing, and reason under open-world conditions. In this thesis, we address these challenges through innovations in data modeling and architectural design. We propose (1) inductive graph alignment prompts for cross-graph transfer, (2) high-frequency-aware contrastive coding for structural robustness, and (3) parameter isolation for continual learning to enable incremental adaptation without catastrophic forgetting. Architecturally, we introduce TransGNN, which combines GNNs with Transformer-based reasoning to bridge local and global semantics, and RAGify, a retrieval-augmented framework that enhances GNN adaptability by moving beyond static graph structures. These contributions establish a foundation for generalizable, adaptive, and deployment-ready graph learning systems. Date: Tuesday, 17 June 2025 Time: 2:00pm - 4:00pm Venue: Room 3494 Lifts 25/26 Committee Members: Dr. Yangqiu Song (Supervisor) Prof. Lei Chen (Chairperson) Dr. May Fung