Graph Neural Networks for Large Dynamic Graphs

Speaker: Yanping Zheng (DATA61-CSIRO)
Host: Dr Sen Wang

Graph Neural Networks for Large Dynamic Graphs

Abstract: In real-world applications such as social networks, financial transactions, and recommender systems, graph-structured data is frequently dynamic, with the nodes and edges of the graph constantly developing dynamically. While Graph Neural Networks (GNNs) have risen as formidable tools for modeling graph-structured data, their adaptation to dynamic graphs introduces distinct challenges. First, any alteration in the graph necessitates a complete relearning of the graph representation matrix, which is expensive and time-consuming. Secondly, even as existing dynamic GNNs are optimized for learning temporal information, they encounter difficulties in scaling to large, evolving graphs. To address these problems, my research focuses on improving the scalability and expressiveness of dynamic graph neural networks.

In my talk, I will first highlight the significance of efficiently computing the graph representation matrix, and then introduce our InstantGNN model. This model incrementally calculates the graph propagation matrix in dynamic graphs, enhancing computational efficiency while ensuring the learning capabilities of GNNs. Next, I will describe the Decoupled Dynamic Graph Neural Network framework. This method supports large dynamic graph learning, where generalized dynamic propagation can effectively support efficient computation on various types of dynamic graphs. Together, these directions help pave a path forward for large dynamic graphs learning.

Speaker Bio: Yanping Zheng is now working as a visiting scientist at DATA61, CSIRO. She is currently a third-year Ph.D. candidate at Gaoling School of Artificial Intelligence, Renmin University of China, advised by Professor Zhewei Wei. Her research focuses on graph learning algorithms. She is particularly interested in efficient algorithms on Graph Neural Networks, Dynamic Graph Representation Learning.