Collaborative Research: Algorithms, Theory, and Validation of Deep Graph Learning with Limited Supervision: A Continuous Perspective

University Of California-Los Angeles, Los Angeles CA

Investigators

Abstract

Graph-structured data is ubiquitous in scientific and artificial intelligence applications, for instance, particle physics, computational chemistry, drug discovery, neural science, recommender systems, robotics, social networks, and knowledge graphs. Graph neural networks (GNNs) have achieved tremendous success in a broad class of graph learning tasks, including graph node classification, graph edge prediction, and graph generation. Nevertheless, there are several bottlenecks of GNNs: 1) In contrast to many deep networks such as convolutional neural networks, it has been noticed that increasing the depth of GNNs results in a severe accuracy degradation, which has been interpreted as over-smoothing in the machine learning community. 2) The performance of GNNs relies heavily on a sufficient number of labeled graph nodes; the prediction of GNNs will become significantly less reliable when less labeled data is available. This research aims to address these challenges by developing new mathematical understanding of GNNs and theoretically-principled algorithms for graph deep learning with less training data. The project will train graduate students and postdoctoral associates through involvement in the research. The project will also integrate the research into teaching to advance data science education. This project aims to develop next-generation continuous-depth GNNs leveraging computational mathematics tools and insights and to advance data-driven scientific simulation using the new GNNs. This project has three interconnected thrusts that revolve around pushing the envelope of theory and practice in graph deep learning with limited supervision using PDE and harmonic analysis tools: 1) developing a new generation of diffusion-based GNNs that are certifiable to learning with deep architectures and less training data; 2) developing a new efficient attention-based approach for learning graph structures from the underlying data accompanied by uncertainty quantification; and 3) application validation in learning-assisted scientific simulation and multi-modal learning and software development. 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.