Learning from Relational Data via Graph Neural Networks
Deep learning has revolutionized various fields of science, including computer vision, natural language processing, and recommendation systems. Though effective deep learning models have been primarily explored for regular-structured data (e.g., images, texts), deep learning for complex non-regular structured data is still challenging. Graph, a prominent example of non-regular data, efficiently describes structural and relational data with wide applications. Adopting deep learning for graphs is undoubtedly important. However, learning from graphs is difficult mainly due to the complex intern patterns posed by arbitrary connectivity. This thesis advances deep learning for graphs using graph neural networks (GNNs). Specifically, we identify three research domains that either implicitly or explicitly present relational data as graphs, and we reveal the great benefits of adopting curated GNNs to these domains: (1) designing GNNs for efficient amortized inference of graphical models. We study Gaussian Process (GP) as a typical example of graphical models. With a proper graph derived from GP prior, we provide an accurate inference network based on GNNs for GP inference. (2) Learning strongly correlated labels among nearby data points. We break the label independence assumption given input features posed by standard GNNs. We adopt our models for spatial data modeling and graph node classification, showing promising results. (3) Learning to solve graph-related combinatorial problems. We develop effective and expressive deep learning models with appropriate training procedures to enable learning GNNs from structured discrete solutions for combinatorial problems. Our model shows superior performance (accuracy and running speed) on graph matching between two graphs. The three domains delivered in this thesis greatly expand the scope of GNNs, further consolidating the possibilities of using deep learning to solve real-world problems.
Please join meeting in Joyce Cummings Center, #180 or via Zoom.
Join Zoom Meeting: https://tufts.zoom.us/j/7084438946
Password: see colloquium email
Dial-in not an option for this event.