Towards Expressive Spectral-Temporal Graph Neural Networks for Time Series Forecasting

被引:0
作者
Jin, Ming [1 ]
Shi, Guangsi [2 ]
Li, Yuan-Fang [3 ]
Xiong, Bo [4 ]
Zhou, Tian [5 ]
Salim, Flora D. [6 ]
Zhao, Liang [7 ]
Wu, Lingfei [8 ]
Wen, Qingsong [9 ]
Pan, Shirui [1 ]
机构
[1] Griffith Univ, Sch Informat & Commun Technol, Southport, Qld 4215, Australia
[2] Monash Univ, Fac Engn, Dept Chem & Biol Engn, Melbourne, Vic 3800, Australia
[3] Monash Univ, Dept Data Sci & AI, Melbourne, Vic 3800, Australia
[4] Univ Stuttgart, Int Max Plank Res Sch Intelligent Syst, D-70174 Stuttgart, Germany
[5] Alibaba Grp, Hangzhou 311121, Peoples R China
[6] Univ New South Wales, Sch Comp Sci & Engn, Sydney, NSW 2033, Australia
[7] Emory Univ, Dept Comp Sci, Atlanta, GA 30322 USA
[8] Anytime AI, New York, NY 10601 USA
[9] Squirrel AI Learning, Bellevue, WA 98004 USA
基金
美国国家科学基金会; 澳大利亚研究理事会;
关键词
Time series analysis; Forecasting; Graph neural networks; Electronic mail; Convolution; Polynomials; Correlation; Australia; Training; Symmetric matrices; Time series forecasting; graph neural networks; spatio-temporal graphs; graph signal processing; deep learning;
D O I
10.1109/TPAMI.2025.3545671
中图分类号
TP18 [人工智能理论];
学科分类号
081104 ; 0812 ; 0835 ; 1405 ;
摘要
Time series forecasting has remained a focal point due to its vital applications in sectors such as energy management and transportation planning. Spectral-temporal graph neural network is a promising abstraction underlying most time series forecasting models that are based on graph neural networks (GNNs). However, more is needed to know about the underpinnings of this branch of methods. In this paper, we establish a theoretical framework that unravels the expressive power of spectral-temporal GNNs. Our results show that linear spectral-temporal GNNs are universal under mild assumptions, and their expressive power is bounded by our extended first-order Weisfeiler-Leman algorithm on discrete-time dynamic graphs. To make our findings useful in practice on valid instantiations, we discuss related constraints in detail and outline a theoretical blueprint for designing spatial and temporal modules in spectral domains. Building on these insights and to demonstrate how powerful spectral-temporal GNNs are based on our framework, we propose a simple instantiation named Temporal Graph Gegenbauer Convolution (TGGC), which significantly outperforms most existing models with only linear components and shows better model efficiency.
引用
收藏
页码:4926 / 4939
页数:14
相关论文
共 55 条
[1]  
[Anonymous], 2020, 8 INT C LEARNING REP
[2]  
Bai L., 2019, PROC INT JOINT C ART
[3]  
Bai SJ, 2018, Arxiv, DOI [arXiv:1803.01271, DOI 10.48550/ARXIV.1803.01271]
[4]   Graph Neural Networks With Convolutional ARMA Filters [J].
Bianchi, Filippo Maria ;
Grattarola, Daniele ;
Livi, Lorenzo ;
Alippi, Cesare .
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, 2022, 44 (07) :3496-3507
[5]  
Boyd S., 2004, Convex optimization, DOI 10.1017/CBO9780511804441
[6]  
Cao D., 2020, P INT C NEUR INF PRO, P17778
[7]  
Choi J, 2022, AAAI CONF ARTIF INTE, P6367
[8]  
Defferrard M, 2016, ADV NEUR IN, V29
[9]   Signed Graph Convolutional Networks [J].
Derr, Tyler ;
Ma, Yao ;
Tang, Jiliang .
2018 IEEE INTERNATIONAL CONFERENCE ON DATA MINING (ICDM), 2018, :929-934
[10]   RELATIVE-ERROR CU R MATRIX DECOMPOSITIONS [J].
Drineas, Petros ;
Mahoney, Michael W. ;
Muthukrishnan, S. .
SIAM JOURNAL ON MATRIX ANALYSIS AND APPLICATIONS, 2008, 30 (02) :844-881