Academic Lecture: Graph Convolutional Neural Networks and Their Applications

Graph Convolutional Neural Networks and Applications

Professor Yan Jingwen

December 3, 2025 (Wednesday) 10:00

This report mainly introduces the latest progress of our team’s research projects on graph convolutional neural networks, specifically the application of graph convolutional networks in traffic flow prediction. To address issues such as time-varying inter-node connectivity and the inability of fixed and empirical graphs to timely and accurately represent information, we first adopted the new LSTSF-GWN model, which combines prior and non-prior knowledge to construct graph structures, effectively capturing complex spatiotemporal correlations; by constructing multiple interpretable graphs, it overcomes the shortcomings of spatial data sparsity in long-term prediction; the model also considers long- and short-term graphs to improve long-term prediction learning capability. Experimental results on two real datasets show that the model’s prediction accuracy outperforms existing deep learning traffic prediction models. Further considering that in practical applications, road traffic speed is affected by various external factors such as weather and social events, future work will introduce some external influencing factors into the model for multi-parameter improvement to further enhance prediction accuracy and robustness.

In addition, this report also introduces the application of graph convolutional neural networks in skeleton-based action recognition, a field that has attracted widespread attention in computer vision. Skeleton-based action recognition faces challenges such as selecting appropriate node features and fully utilizing semantic information in the skeleton. To address these issues, this paper proposes two solutions and network models. First, a Semantic Decomposition Graph Convolutional Network (SD-GCN) is proposed, fully utilizing skeleton semantic information by inputting fused 2D and 3D skeletons as node features into the network to improve classification accuracy. Second, to further improve the classification accuracy of SD-GCN, a Multi-Relationship Graph Attention Network (MRGAN) is proposed. Relationships are treated as subgraphs, and node correlations are accurately modeled through weighted fusion of subgraph information. The multi-relationship graph attention mechanism focuses on key action parts, improving accuracy and robustness. Experimental results on four large datasets demonstrate excellent performance.

Yan Jingwen, male, PhD, Professor, doctoral supervisor in Communication and Systems at Xiamen University and Fundamental Mathematics at Shantou University; Deputy Director of Key Lab of Digital Signal and Image Processing of Guangdong Province; Council Member of China Society of Image and Graphics; Deputy Director of the Science Popularization and Education Committee of China Society of Image and Graphics; Executive Council Member of Guangdong Society of Image and Graphics; Expert of Guangdong Security Association. Key introduced talent in Xiamen in 2006 and in Shantou in 2008; recipient of Xiamen University’s first batch of Level III positions and Level III professorship.

July 1987, Bachelor’s degree in Engineering, Department of Electronic Engineering, Jilin Polytechnical University (now Jilin University); July 1992, Master’s degree in Science (Cartography and Remote Sensing), Northeast Institute of Geography and Agroecology, Chinese Academy of Sciences; December 1997, PhD in Science (Optics), Changchun Institute of Optics, Fine Mechanics and Physics, Chinese Academy of Sciences. September 2002–August 2003, Visiting Scholar at ParisTech (Department of Signal and Image Processing, Télécom Paris, engaged in SAR image processing research); August 2010–November 2010, Visiting Professor at the H. Milton School of Industrial and Systems Engineering, Georgia Institute of Technology (GaTech), collaborated with Professor Huo Xiaoming on astronomical image analysis and data mining; August 1998–September 2006, Lecturer, Associate Professor (December 1999), and promoted to Professor (December 2003) at the Department of Electronic Engineering, Xiamen University; September 2006–present, Professor at the Department of Electronic Engineering, College of Engineering, Shantou University.

As first author and corresponding author, published nearly 100 papers in journals such as Acta Electronica Sinica, Acta Automatica Sinica, Acta Optica Sinica, Chinese Optics Letters, Optics and Precision Engineering, Journal of Image and Graphics, and international/domestic conferences, including 20 SCI-indexed, 30 EI-indexed, and 10 ISTP-indexed papers. Led and completed 3 projects of the National Natural Science Foundation of China, more than 10 provincial/ministerial projects, and several horizontal projects; authored, edited, and translated 5 books.

Prepared by: Ren Luyang

Typeset by: Chen Shifa

First Review and Proofreading: Ren Luyang

Second Review and Proofreading: Ma Jiang

Third Review and Proofreading: Zheng Chun