报告人: Yuzhou Chen
报告主题: Topological Representation Learning on Graphs: Methods and Applications
In the last few years, Geometric Deep Learning (GDL), e.g., Graph Neural Networks (GNNs), have emerged as a powerful alternative to more conventional deep learning (DL), machine learning (ML), and statistical models from clustering/anomaly detection to node classification/link prediction to time-series forecasting. Despite their proven success, GNNs tend to be limited in their ability to simultaneously infer latent temporal relations and encode higher order interactions among entities. Our works tackle these limitations across a spectrum of higher-order network analysis, from topological data analysis to simplicial complexes. In our works, (i) we develop a novel time conditioned topological representation, and make the first step on a path of bridging the two emerging directions, namely, time-aware GDL with time-conditioned topological representations of complex dynamic networks, and (ii) we bridge the recently emerging concepts of convolutional architectures on witness complexes with topological signal processing on graphs with applications on bioinformatics and social networks classification. This talk will highlight two projects that epitomize these methodologies. First, I will present our topology-based spatio-temporal GDL model TAMP-S2GCNets --- the first effort to bridge topology-based GDL model with time-aware multiparameter persistent homology representations of the data in learning complex multivariate spatiotemporal processes. I will discuss our works on traffic flow forecasting, Ethereum blockchain price prediction, and COVID-19 hospitalizations forecasting, and also show substantial computational gains and high utility of the proposed time-conditioned topological descriptors for encoding the time-conditioned knowledge. Second, I will present a novel topological pooling method Wit-TopoPool for graph classification tasks. The corresponding GNN-based topological pooling layer and witness complex-based topological layer in this model can learn rich discriminative topological characteristics of the graph as well as to extract essential information from node features. Finally, I will conclude with a number of exciting and important applications and directions for future work.
Yuzhou Chen is Assistant Professor in the Department of Computer and Information Sciences at Temple University. Prior to joining Temple University, he was a Postdoctoral scholar in the Department of Electrical and Computer Engineering at Princeton University. He received his Ph.D. in Statistics at Southern Methodist University in 2021. Before that, he was a research fellow at Lawrence Berkeley National Laboratory, National Renewable Energy Laboratory and INRIA, respectively. He received his M.S. degree from the University of Texas at Dallas. His main research interests are in geometric deep learning, topological data analysis, knowledge discovery in time series, blockchain, and power systems. He has published several papers in machine learning, data mining top conferences, including ICML, ICLR, NeurIPS, KDD, AAAI, ICDM, etc. He was the recipient of the 2022 and 2021 Best Paper Award of the Section for Statistics in Defense and National Security (SDNS) of the American Statistical Association (ASA) and the 2021 Chateaubriand Fellowship.