Associate Professor Jingrui He has been awarded a two-year, $149,921 grant from the National Science Foundation (NSF) to improve the performance of deep learning models. For her project, "Weakly Supervised Graph Neural Networks," she will focus on the lack of labeled data in Graph Neural Networks (GNNs), a deep learning method designed to perform inference on data described by graphs.
A graph is a structured way to represent data, with nodes representing entities and edges representing the relationships between these entities. GNNs provide an easy way to conduct node-level, edge-level, and graph-level prediction via machine learning. However, they usually require a large amount of label information to train the model parameters. According to He, the lack of labeled data in graphs can render many existing deep learning models ineffective in achieving the desired performance. Her new project involves a work-around so that GNNs can use unlabeled data and other relevant information.
"For example, in fraud detection, the number of known fraudulent transactions is usually very small compared to the total number of transactions, hence the lack of labeled data. Most existing GNN models tend to suffer from such label scarcity. In my new project, we aim to address this issue by leveraging weak supervision or additional information (besides the limited label information), such as labeled data from other related applications and/or access to a domain expert, in order to compensate for the lack of labeled data," said He.
In addition to fraud detection, areas such as agriculture and cancer diagnosis could also benefit from this research. He’s project will lead to a suite of new models, algorithms, and theories for constructing high-performing GNNs with weak supervision, and for understanding the benefits of weak supervision with respect to the model generalization performance and sample complexity.
He's general research theme is to design, build, and test a suite of automated and semi-automated methods to explore, understand, characterize, and predict real-world data by means of statistical machine learning. She received her PhD in machine learning from Carnegie Mellon University.