December 16, 2019 –
Title: End-to-End Structure-Aware Convolutional Networks on Graphs
Ph.D. Candidate: Chao Shang
Major Advisor: Dr. Jinbo Bi
Associate Advisors: Dr. Alexander Russell and Dr. Yufeng Wu
Review Committee Members: Dr. Derek Aguiar and Dr. Caiwen Ding
Date/Time: Monday, December 16, 2019 10:00-11:00 AM
Location: ITEB 336
Convolutional Neural Networks (CNNs) are powerful tools to model data of a grid-like structure, such as image, video, and speech. However, a broad range of scientific problems generate data that naturally lie in irregular grids with non-Euclidean metrics, such as molecular graphs and knowledge graphs. The generalization of CNNs to non-Euclidean structured data such as graphs is not straightforward. The classical convolutions cannot be applied directly to graphs, due to the lack of global parameterization, a common system of coordinates, and shift-invariance properties.
In this dissertation, we propose several structure-aware convolutional network models to calculate graph convolutions efficiently over both small-scale and large-scale graphs. The proposed networks can be trained using an end-to-end training method with a stochastic gradient algorithm back propagating over all network components rather than a stage-wise training scheme where the different components are tuned separately. The first part of the proposal focuses on large-scale knowledge graphs. We present a novel approach that learns the graph connectivity structure to infer new connectivities in a knowledge graph and maps an input knowledge graph to a more complete graph. We call this approach "end-to-end structure-aware (aware of the edge structures) convolutional network". This model not only utilizes the node (or entity) attributes and edge relations in a knowledge graph but also preserves the so-called translational property between entities and relations. In the second part of the proposal, we extend the convolution operation to small-scale hydrogen-depleted molecular graphs. Unlike the first model that learns from a single massive graph, this method learns from a massive amount of small graphs. We propose a consistent edge-aware (aware of the edge consistency in molecular structures) convolutional neural network to predict a molecule's properties based on its molecular graph. This model exploits the general consistency of the bond energies and bond lengths across various molecular graphs. Extensive experiments demonstrate the substantial advantages of the proposed techniques in both knowledge graph completion and molecular quantitative structure-activity relationship prediction.