Data Stream Management: Efficient T-GNN Training over Large-Scale Dynamic Graphs

The Hong Kong University of Science and Technology (Guangzhou)

Data Science and Analytics Thrust

PhD Thesis Proposal Examination

By Mr GAO Shihong

Abstract

Temporal Graph Neural Networks (T-GNNs) have become the de facto solution for representation learning on dynamic graphs, enabling state-of-the-art performance on tasks such as temporal link prediction and recommendation. However, existing T-GNN training pipelines suffer from scalability issues due to ill-suited batching and high input data loading costs, which severely limit their efficiency on large-scale graphs. This thesis proposal addresses both these bottlenecks with two complementary system prototypes. First, we propose ETC, a generic framework that introduces a theoretically grounded batch splitting algorithm and a three-step deduplication policy to improve computation throughput and reduce I/O overhead. Second, we present SIMPLE, a dynamic data placement system that maintains a GPU buffer for frequently accessed inputs, optimizing data reuse through an interval selection algorithm with approximation guarantees. Together, ETC and SIMPLE significantly accelerate T-GNN training, achieving up to 62.4× speedup over state-of-theart baselines while preserving model accuracy, as demonstrated by extensive experiments on real-world datasets.

TPE Committee

Chair of Committee: Prof. ZHOU, Xiaofang(Online)

Prime Supervisor: Prof. YANG, Can (Online)

Co-Supervisor: Prof. CHEN, Lei  

Examiner: Prof. ZHANG, Yongqi