CAREER: Advancing Graph Neural Networks via Graph Compression

Many real-world domains can be represented as graphs. For instance, a roadmap can be represented as a graph where cities are the nodes and roads connecting cities are the edges connecting the nodes. A Graph Neural Network (GNN) is a deep learning model designed to operate on graphs. There are many domains that are amenable to such graph representations, including transportation, biomedical, social, and security domains. As a result, GNNs are the subject of increasing interest by researchers and practitioners. However, GNNs need to cope with the complexity of real-world networks, such as their diverse connectivity patterns, and dynamic nature. The goal of this CAREER proposal is to develop novel GNN models that create a smaller compressed graph preserving the desired structural information of large graphs. This will enable GNNs to be deployed in more real-world applications. Compressing graphs will also achieve several benefits, including significant time and memory space reduction, and improved data privacy. The project will integrate research with education by organizing several activities including outreach to high school and undergraduate students, underrepresented groups, and HBCUs to attract women and minorities to the computing field. The objective of this project is to develop structure-aware graph compression methods by exploring the critical topological properties of graphs and build tailored architectures by integrating proposed compression methods into GNN models. To achieve the proposed project objective, this research aims to systematically pursue the following interconnected research tasks: 1) create novel graph compression techniques that tackle the oversmoothing problem common to GNNs, 2) develop novel graph compression methods capable of capturing local, global, and higher-order structural information encompassed within the entire graph to facilitate pooling operations, 3) devise compression methods that effectively capture and compress the shared structures present in multidimensional networks, thus incorporating them directly into the learning process and eliminating redundancy. Furthermore, the project will evaluate the performance of the developed models by applying them to drug-disease matching and event-detection problems with collaborating experts from different domains. 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.