How Powerful are Spectral Graph Neural Networks?

The LOGS (Graph Learning Seminar) organized a talk on July 9, 2022 at 10:00 Beijing Time, featuring Prof. Zhang Muhan from Peking University.

Title: “How Powerful are Spectral Graph Neural Networks?” (谱图神经网络有多强大？) and hosted by Zhou Min from Huawei Noah’s Ark Lab.

Abstract: Spectral GNNs are graph signal filter‑based networks widely used for node tasks. Many popular GNNs (ChebyNet, GCN, BernNet) belong to this family, yet their expressive power has been under‑studied. The talk simplifies spectral GNNs to Linear GNNs, proves a universal approximation theorem under mild conditions, relates their expressiveness to the Weisfeiler‑Lehman test, and proposes a new spectral GNN called JacobiConv that uses a Jacobi basis and achieves state‑of‑the‑art performance with only linear layers.

Reference: Xiyuan Wang, Muhan Zhang, “How Powerful are Spectral Graph Neural Networks,” ICML 2022. arXiv: https://arxiv.org/abs/2205.11172v2. Code: https://github.com/GraphPKU/JacobiConv

Speaker bio: Dr. Zhang Muhan is an assistant professor at Peking University AI Institute, researcher at Beijing General AI Institute, and a National Excellent Young Scientist (overseas) awardee. He earned his Ph.D. from Washington University in St. Louis, worked as a Facebook AI Research scientist, and has published 15 papers at top ML conferences, contributing classic algorithms such as SortPooling, SEAL, and IGMC. He serves as area chair for major conferences and reviewer for leading journals.

Participants can reply “logs” to the public account to obtain the live stream link.