Towards Large-Scale Graph Neural Networks: A Survey on Sampling-Based Methods

PhD Qualifying Examination


Title: "Towards Large-Scale Graph Neural Networks: A Survey on Sampling-Based 
Methods"

by

Miss Jingshu Peng


Abstract:

Graph data has been prevalent with its ability to model many real-life 
applications woven by huge and complex networks of relations and interactions, 
such as the social networks and even the universe. Recently, graph neural 
networks (GNNs) have come into the spotlight in their great success to model 
the tangled relations in graph data, as a powerful machine learning tool. In 
GNNs, the neural networks are built up on the basis of the original graph 
structure, leveraging both feature and topology information iteratively from 
the neighborhood to capture the complex relationships and interdependencies 
between entities in the graphs. However, the major challenge that limits the 
adoption and deployment of GNNs to large-scale graphs in the real world, 
despite their promising performance, lies in the inability to utilize all the 
data in finite time and the scalability of the algorithm itself. Due to its 
iterative neighborhood aggregation nature, the training of GNNs intrinsically 
gives rise to high computational cost and massive storage requirements. 
Tremendous efforts have been made towards sampling-based mini-batch GNN 
training to mitigate the memory requirement. Therefore, various sampling 
strategies have been proposed, aiming to improve the training efficiency and 
scalability of current GNN models over large-scale graphs, such as node-wise 
sampling, layer-wise importance sampling, and the fundamentally different 
subgraph sampling. In this paper, we present a comprehensive review of the 
sampling-based GNNs. First of all, we overview the basics and fundamentals of 
GNNs. Then, we introduce, categorize and analyze representative GNN sampling 
strategies that lay the foundation, to get a grasp about how this line of works 
in the GNN field evolves. Further, we highlight a detailed comparison for 
different sampling techniques to understand their strengths and weaknesses. 
Finally, we discuss possible future directions for the sampling approach.


Date:			Tuesday, 7 December 2021

Time:			8:30am - 10:30am

Zoom Meeting:
https://hkust.zoom.us/j/95543240942?pwd=a3VJeU02NmN5WnVJcGcrVnU3ZUdpUT09

Committee Members:	Prof. Lei Chen (Supervisor)
 			Prof. Xiaofang Zhou (Chairperson)
 			Prof. Qiong Luo
 			Prof. Ke Yi


**** ALL are Welcome ****