III: SMALL: Graph Contrastive Learning for Few-Shot Node Classification

Arizona State University, Scottsdale AZ

Investigators

Abstract

A graph is a data structure consisting of nodes and edges. Graph data is the data associated with nodes and edges in a graph. Graph data is huge and is widely present in many real-world applications. Social media data is a typical example of graph data in which users are nodes and their relationships are edges. Since users have different profiles, they can form disparate relationships (i.e., edges) amongst themselves. When a dataset is large, annotating or labeling it with ground truth is time consuming and labor intensive. A pressing need for machine learning and data mining to effectively deal with big data like graph data is to address the labeled data scarcity problem. When we can only label a miniscule amount of data, can we learn well from big graph data? Graph few-shot node classification, in which learning can occur when only a small amount of data are labeled, is one such problem for which researchers strive to find novel solutions. In such a problem, training data can vary in the training phase depending on the availability of labeled nodes - with labels, weak labels, or no labels. To address such unprecedented challenges, this project aims to develop new approaches. The proposed research will train students to perform independent research, conduct scientific experiments, and publish technical results to nurture science and engineering researchers. Students will be exposed to the core techniques of real-world problems with graph data and machine learning. The impact of this work will also extend to critical thinking of dominant approaches, understanding the essence of difficult problems such as graph few-shot node classification, and exploring simple and effective solutions considering real-world scenarios. Episodic meta-learning is currently the dominant approach that has been shown to be effective for supervised few-shot node classification. This project questions the necessity of this meta-learning approach and elaborates the need for a novel graph contrastive learning approach to few-shot node classification to handle supervised few-shot node classification and more challenging and realistic cases where only weak or no supervision information is available during training. This project investigates an alternative approach - graph contrastive learning in search of a general learning framework for the challenging problem of few-shot node classification and to handle the cases with noisy or no labels during training by examining fundamental research issues and developing new algorithms for supervised, weakly supervised, and self-supervised few-shot node classification. Related work is reviewed, preliminary studies related to each research task are presented, and innovative research tasks are proposed to develop original and systematic solutions. With the proven track record in graph learning and insights gained in the preliminary studies for each proposed research task by the PI’s team, this project is envisioned as laying a solid foundation for graph contrastive learning for few-shot node classification and paving the way to advance the frontier of learning graph data with noisy or no labels during training. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.