More about HKUST
Learning in Flux: Continual and Efficient Representation Learning on Dynamic Graphs
The Hong Kong University of Science and Technology Department of Computer Science and Engineering PhD Thesis Defence Title: "Learning in Flux: Continual and Efficient Representation Learning on Dynamic Graphs" By Mr. Peiyan ZHANG Abstract: The ability to learn and adapt within continuous data streams is a hallmark of true intelligence, yet it remains a grand challenge for modern artificial intelligence. Dynamic graphs offer a natural representation for these systems, yet prevailing graph learning models often fail to capture the temporal fidelity required for true, real-time adaptation. This creates a fundamental gap between existing analytical capabilities and the fluid reality of the environments we aim to model. This thesis confronts this challenge by proposing and validating a new methodology centered on the Dual-Fidelity framework. We posit that achieving genuine living models requires a simultaneous pursuit of fidelity along two orthogonal dimensions. The first is Data-Level Representation, where time should be treated as an intrinsic driver of state evolution rather than a mere external feature. The second is the Process-Level Paradigm, which requires proactive, continual learning mechanisms that explicitly combat catastrophic forgetting. To materialize this framework, this thesis presents a sequence of technical contributions addressing each dimension. For the data-level dimension, we develop Graph Nested GRU-ODE (GNG-ODE), an architecture integrating Neural Ordinary Differential Equations into dynamic graph-based recommendation. It models user and item states as continuous trajectories governed by the precise time elapsed between events. For the process-level dimension, we design the Parameter-Isolation Graph Neural Network (PI-GNN). It introduces a mechanism to isolate stable and active parameters, which enables incremental adaptation to new data while mitigating catastrophic forgetting. Finally, to ensure this high-fidelity methodology is not merely theoretical but practical for large-scale industrial settings, the thesis culminates in Graph Prompt Tuning for Recommender Systems (GPT4Rec). This paradigm resolves the critical tension between adaptation and efficiency by decoupling a large, frozen knowledge base from a small set of trainable prompts. It demonstrates that pursuing high fidelity is not only effective but also computationally feasible in demanding real-world environments. In synthesis, this work provides a new analytical framework for dynamic graph learning, which offers a validated technical roadmap for achieving higher fidelity in both data representation and learning processes, and presents a scalable solution for its practical deployment. The research contributes to the development of AI systems capable of real-time reasoning and adaptation in evolving worlds. Date: Friday, 19 September 2025 Time: 3:00pm - 5:00pm Venue: Room 5501 Lifts 25/26 Chairman: Prof. Daniel PEREZ PALOMAR (ECE) Committee Members: Dr. Yangqiu SONG (Supervisor) Dr. Shuai WANG Dr. Dan XU Prof. Jun ZHANG (ECE) Dr. Xiangyu ZHAO (CityU)