EAGER: Weakly Supervised Graph Neural Networks

Graph Neural Networks have proven to be a powerful tool for harnessing graph data, which is widely used for representing rich relational information in multiple areas. However, the performance of graph neural networks largely depends on the amount of labeled data, and thus can be significantly affected by the label scarcity caused by the expensive and time-consuming annotation process, which is common among many high-impact applications, such as fraud detection, agriculture, cancer diagnosis.

This project focuses on building high-performing graph neural networks in the presence of label scarcity. In particular, the developed techniques advance state-of-the-art by systematically leveraging weak supervision in the form of relevant source graphs and access to a labeling oracle with a limited budget. This project generates a suite of new models, algorithms and theories for constructing high-performing graph neural networks with weak supervision, and for understanding the benefits of weak supervision from a theoretical perspective.

It advances state-of-the-practice for graph neural networks by significantly reducing the need for large amount of labeled data. This project involves students at various levels, especially those from under-represented groups. The research outcomes from this project will be disseminated at relevant conferences and journals in computer science.

This project consists of two complementary research thrusts, focusing on the pre-training stage and the fine-tuning stage of the model construction process for graph neural networks respectively. For the pre-training stage, given the rich information from relevant source graphs, this project develops techniques to leverage such information via cross-graph domain adaptation, in order to obtain effective representation of the target graph at various granularities; for the fine-tuning stage, given a limited budget for querying an oracle, this project develops techniques to select the most informative nodes/edges/subgraphs based on the training dynamics of graph neural networks, such that this additional label information can maximally improve the model performance.

Furthermore, this project establishes new theoretical results regarding the benefits of weak supervision, such as the impact of source graphs on the model generalization performance and the reduction of the sample complexity due to active learning with graph neural networks.

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.