Date of Completion
Structure-Aware Learning; Graph Neural Networks
Field of Study
Computer Science and Engineering
Doctor of Philosophy
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 knowledge graphs, molecular graphs, and traffic networks. 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 neural network models to calculate graph convolutions efficiently over both small-scale and large-scale graphs. The proposed networks can be trained by an end-to-end training method where a stochastic gradient descent algorithm back-propagates over all network components rather than a stage-wise training scheme where the different components are tuned separately. These models are built for exploring the structure information to improve many prediction tasks: (1) The first part of the dissertation focuses on a large-scale knowledge graph. Our new approach learns the graph connectivity structure so it can infer new edges in a knowledge graph and grow an input knowledge graph to be more complete. 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. (2) The second part extends the convolution operation to small-scale hydrogen-depleted molecular graphs. Unlike the first model that learns from a single graph of massive size, this method learns a novel graph-based model from a massive amount of small graphs. We propose a consistent edge-aware graph convolutional network that determines consistent edge attentions to the same type of edges appearing in different molecular graphs and predicts a molecule's properties. This model exploits the general consistency of the bond energies and bond lengths across various molecular graphs. (3) The third part focuses on the exploration of the correlation and causation among multivariate time series in a graph where a node represents a time series variable and an edge indicates if a time series causes another time series. When the connectivity (edge) structure is not available or incomplete, we propose a discrete structure learning method that learns the hidden graph structure simultaneously when constructing the predictive model for the time series data. Extensive experiments demonstrate the advantages of the proposed techniques over the state of the art in knowledge graph completion, molecular quantitative structure-activity relationship prediction, and multivariate time series forecasting.
Shang, Chao, "End-to-End Structure-Aware Convolutional Networks on Graphs" (2020). Doctoral Dissertations. 2555.