"For me, this NSF award gives me great confidence and provides necessary funding resources for my research on graph neural networks," Gao said.
Project Abstract: Many data in the real world can be naturally represented by graphs, such as social networks, citation networks, chemical compounds, and biological networks. The graph structure is an effective way to express relationships among a collection of items. Using graphs to represent these data, we can obtain an advanced understanding of the complex data structures embedded in the raw data. Graph mining using machine learning methods yields many exciting discoveries in various fields. Existing graph deep learning methods, such as graph convolutional networks (GNN), conduct local information aggregation with local structural operations. They usually stack multiple graph convolutional layers to enable a larger receptive field, which is straightforward but can result in several issues, including over-fitting and over-smoothing. These issues are critical to graph neural networks but are not well investigated. The objective of this project is to develop advanced graph learning methods without going deep and complex. This project also facilitates integrating graph learning algorithms into existing curricula of Machine Learning courses at both undergraduate and graduate levels.
Specifically, this project focuses on how to enlarge receptive fields without going deep effectively. This project develops a set of advanced convolution, pooling, and un-pooling operations on graphs. The advanced graph convolution layer utilizes teleport functions to select highly relevant nodes at the global scope. In this layer, a teleport function computes relevance between the center node and other nodes beyond the local neighborhood. The nodes with particular relevance are teleported for the center node, enabling the center node to gather information from a broader neighborhood. The new graph pooling layer uses an attention operation to produce a better-connected coarsened graph with more graph topology information. In particular, it leverages an attention operator to generate ranking scores for node selection, capturing the graph connectivity information. The proposed graph un-pooling layer utilizes an attention operator to initialize restored nodes. Altogether, this project provides a systematic GNN study on graph feature and structure learning and is expected to advance GNNs development significantly.