Graph Neural Networks: A Feature and Structure Learning Approach

Monday, March 2, 2020 - 10:15am to 11:15am
Storey Innovation Center (Room 2277)

In the real world, many data are naturally represented as graph data such as social networks. Deep learning methods have been very successful in various fields such as computer vision and natural language processing. However, developing deep learning methods on graph data is challenging due to the lack of locality information. In this talk, I will present my work on developing deep learning methods on graph data. My work addresses this challenge and significantly advances feature learning and structure learning on graphs in both accuracy and efficiency. Specifically, I will introduce our proposed learnable graph convolution layer and hard graph attention layer, which enables fully learnable convolution and hard attention operations on graph data while saving computational resources. Then I will discuss our developed efficient and effective graph pooling operators that significantly advance state-of-the-art performance. Besides layer-wise methods, I will talk about the first encoder-decoder network architecture on graph data. This series of research works result in a series of publications in top-tier journals and conferences.

Bio:
Hongyang Gao is a Ph.D. Candidate in the Department of Computer Science & Engineering at Texas A&M University in College Station, Texas. His primary research interests are machine learning and artificial intelligence with a special focus on deep learning. In particular, he mainly pays attention to the performance and efficiency of deep learning methods with applications to various data types like graphs. His research work has been recognized with a series of publications in top-tier journals and conferences. Before his Ph.D. work, Hongyang received his M.S. in Computer Science from Tsinghua University in 2012 and his B.S. from Peking University in 2009.

Monday 3/2/20 at 10:15am