Graph Neural Networks: From Theory to Applications

Date&Time:13:30 – 15:30, December 28, 2021   Location: Meeting Room 1

Intro: Graphs are flexible mathematical structures modeling pairwise relations between data entities, such as brain networks, social networks, transportation networks, knowledge graphs, irregular geometric data, etc. Graph Neural Networks (GNNs) are generalizations of CNNs to graph-structured data, which have been attracting increasing attention due to the great expressive power. Thanks to its convincing performance and high interpretability, GNNs have been widely employed to represent various kinds of multimedia data, such as social networks, 3D point clouds, natural languages, etc. This symposium will invite four speakers to share their insights and cutting-edge results in Graph Neural Networks from theory to applications.


Organizer: Wei Hu (Peking University) and Yue Gao (Tsinghua University)




Spatio-temporal Graph Networks for Multi-gent Systems

Siheng Chen


Cross-domain Feature Learning for 3D Model Retrieval

Weizhi Nie


Hierarchical Graph Convolutional Networks for Multiscale Representation Learning

Wenrui Dai


Two Sides of Graph Neural Networks: Characteristics and Problems

Xiao Wang


Title: Spatio-temporal Graph Networks for Multi-gent Systems

Speaker: Siheng Chen,Shanghai Jiao Tong University

Abstract: Irregular spatio-temporal data are irregularly-sampled data that have both spatial and temporal information. Irregular spatio-temporal data processing and analysis has become the core technique in real-world systems, such as autonomous driving and surveillance systems. It is also considered as one of the most cutting-edge research topics in both signal processing and machine learning. One emerging approach to handle irregular spatio-temporal data is to model them on spatio-temporal graphs and leverage graph neural networks to extract information. In this talk, we introduce our recent research progress from three perspectives. From an application perspective, we discuss collaborative prediction in a multi-agent system, which aims to forecast the future states of each agent by considering the social influence from other agents; from an algorithmic perspective, we introduce spatio-temporal graph neural networks, which expands deep learning techniques to the spatio-temporal graph domain; and from a theoretical perspective, we propose a new mathematical framework, spatio-temporal graph scattering transform, which strategically combines the advantages of both graph signal processing and graph neural networks, aiming to achieve theoretical interpretability and provide convincing practical performances.

Biography: Siheng Chen is an associate professor at Shanghai Jiao Tong University. Before that, he was a research scientist at Mitsubishi Electric Research Laboratories (MERL) and an autonomy engineer at Uber Advanced Technologies Group, working on the perception and prediction systems of self-driving cars. Before joining an industry, he was a postdoctoral research associate at Carnegie Mellon University. He received the doctorate in Electrical and Computer Engineering from Carnegie Mellon University in 2016, where he also received two masters degrees in Electrical and Computer Engineering and Machine Learning, respectively. He received his bachelor’s degree in Electronics Engineering in 2011 from Beijing Institute of Technology, China. His paper “Discrete signal processing on graphs: Sampling theory” won the 2018 IEEE Signal Processing Society Young Author Best Paper Award. His coauthored paper received 2020 ASME SHM/NDE Best Paper Runner-Up Award and Best Student Paper Award at IEEE GlobalSIP 2018. He was selected in National Young Talent Plan in 2019. He serves as an Associate Editor for IEEE Transactions on Signal and Information Processing over Networks and session chairs for the International Conference on Acoustics, Speech, & Signal Processing (​ICASSP) 2020, 2021. His research mainly focuses on graph-structured data science and its applications to autonomous systems.


Title: Cross-domain Feature Learning for 3D Model Retrieval 

Speaker: Weizhi Nie,Tianjin University

Abstract: With the advanced development of digitalization techniques and computer vision, 3D shapes are commonly used in our everyday lives, such as in computer-aided design, medical diagnosis, bioinformatics, 3D printing, medical imaging, and digital entertainment. In particular, applications in virtual and augmented reality create a demand for rapid creation and easy access to large sets of 3D shapes in recent years. It is reasonable to utilize some references to obtain similar 3D shapes and accelerate the secondary development. This report introduces the technologies of 3D model retrieval and reconstruction based on cross-domain feature learning methods. Some classic approaches will be illustrated in this report. I also will introduce some new work from our team.

Biography: Dr. Weizhi Nie is currently an Associate Professor at Tianjin University, graduated from Tianjin University in 2015. His research focuses on machine learning, information retrieval, data mining, and computer vision. He has published 50+ peer-reviewed papers (CCF-A) in top venues in multimedia and CV. He received Best Paper awards in ChinaMM 2020. He received Tianjin Science and Technology Progress Award in 2019 and 2020.


Title: Hierarchical Graph Convolutional Networks for Multiscale Representation Learning

Speaker: Wenrui Dai,Shanghai Jiao Tong University

Abstract: Graph neural networks have emerged as a popular and powerful tool for learning multiscale representation of graph data. In complement to graph convolution operators, graph pooling is crucial for extracting hierarchical representation of data in graph neural networks. However, most recent graph pooling methods still fail to efficiently exploit the geometry of graph data. In this talk, I will present our recent investigations on graph pooling that consider both graph signals and topology. I will first introduce a parameter-free pooling operator that permits to retain the most informative features in arbitrary graphs. To preserve the informative nodes dominantly characterizing the graph signals, a criterion is developed to evaluate the amount of information of each node given its neighbors and is demonstrated to relate to neighborhood conditional entropy in theory. Furthermore, node affinity is computed by harmonizing the kernel representation of topology information and node features. A structure-aware kernel representation is presented to explicitly exploit advanced topological information without eigen-decomposition of the graph Laplacian. Experimental results demonstrate that these graph pooling operators achieve superior or competitive performance in graph classification on a collection of public graph benchmark datasets and super-pixel induced image datasets.

Biography: Wenrui Dai received B.S., M.S., and Ph.D. degree in Electronic Engineering from Shanghai Jiao Tong University, Shanghai, China in 2005, 2008, and 2014. He is currently an associate professor at the Department of Computer Science and Engineering, Shanghai Jiao Tong University (SJTU). Before joining SJTU, he was with the faculty of the University of Texas Health Science Center at Houston from 2018 to 2019. He was a postdoctoral scholar with the Department of Biomedical Informatics, University of California, San Diego from 2015 to 2018 and a postdoctoral scholar with the Department of Computer Science and Engineering, SJTU from 2014 to 2015. His research interests include image/signal processing, learning-based image/video coding and predictive modeling. He has published 60 papers in prestigious journals like IEEE Transactions on Pattern Analysis and Machine Intelligence, IEEE Transactions on Image Processing, IEEE Transactions on Signal Processing and conferences like ICML and CVPR.


Title: Two Sides of Graph Neural Networks: Characteristics and Problems

Speaker: Xiao Wang,Beijing University of Posts and Telecommunications

Abstract: Graph neural networks (GNNs), as one of the most important techniques dealing with graph data in both of the academic and industrial areas, have been a new wave in current deep learning. This talk will systematically rethink the two sides of GNNs: the characteristics and the resulting problems. The superior performance of GNNs usually depends on the high-quality input graph and the low-frequency information extraction, etc, which are actually easily hindered in real-world scenarios. This talk will introduce the recent advances in graph structure learning, multi-channel GNNs, frequency-adaptive GNNs, and the final unified framework of GNNs. This may bring new insight on understanding GNNs, and provides more powerful and comprehensive representation ability for GNNs.

Biography: Dr. Xiao Wang is currently an Assistant Professor with Beijing University of Posts and Telecommunications. Before that, he was a postdoc in the Department of Computer Science and Technology at Tsinghua University. He got his Ph.D. at Tianjin University, China. He was also a joint-training Ph.D. at Washington University in St. Louis, USA. His current research interests include data mining and machine learning. He has published 50+ peer reviewed papers (CCF-A) in top venues in AI, where one paper is ESI highly cited, one paper got WWW 2021 best paper awards nomination, and three papers are selected as the most influential papers in AAAI 2017/WWW 2019, 2020 by PaperDigest. He has presided the project of National Nature Science Foundation of China (NSFC) and CCF-Tencent Open Research Fund. He also serves as SPC/PC in ACM SIGKDD, AAAI, IJCAI, WWW, etc.