CVPR 24' Tutorial - Unifying Spectral and Spatial Graph Neural Networks

Intro

Over recent years, Graph Neural Networks (GNNs) have garnered significant attention. However, the proliferation of diverse GNN models, underpinned by various theoretical approaches, complicates the process of model selection, as they are not readily comprehensible within a uniform framework. Specifically, early GNNs were implemented using spectral theory, while others were developed based on spatial theory . This divergence between spectral and spatial methodologies renders direct comparisons challenging. Moreover, the multitude of models within each domain further complicates the evaluation of their respective strengths and weaknesses.

In this half-day tutorial, we examine the state-of-the-art in GNNs and introduce a comprehensive framework that bridges the spatial and spectral domains, elucidating their complex interrelationship. This emphasis on a comprehensive framework enhances our understanding of GNN operations. The tutorial’s objective is to explore the interplay between key paradigms, such as spatial and spectral-based methods, through a synthesis of spectral graph theory and approximation theory. We provide an in-depth analysis of the latest research developments in GNNs in this tutorial, including discussions on emerging issues like over-smoothing. A range of well-established GNN models will be utilized to illustrate the universality of our proposed framework.

Materials/Resources to be distributed

• CVPR 2024 tutorial official page
• Slides version for CVPR 24’
• Major reference paper in ACM Computing Survey
• A collection of related sources at GitHub.

Presenters, Time and Location

Room 440-441, 2pm, June 18, 2024.

Tentitative Program overview

The total duration comprises 150 minutes allocated for the tutorial, 15 minutes for a break, and 15 min for Q&A.

1, Background: Spectral and Spatial Graph Neural Networks (20 min)

Graph Neural Networks (GNNs) can be bifurcated into two primary categories based on their underlying computational mechanisms: spectral and spatial approaches. Spectral GNNs, grounded in graph signal processing, hinge upon the eigen-decomposition of the graph Laplacian, with foundational contributions such as ChebNet. These networks furnish a theoretically robust framework for articulating convolution operations on graphs, albeit potentially exacting in terms of computational resources. In contrast, spatial GNNs operate within the vertex domain, aggregating information from neighboring nodes, exemplified by GraphSAGE and GCN. This tutorial provides an extensive overview of the recent advancements in both spectral- and spatial-based GNNs, and embarks on a detailed analysis of the evolutionary trajectories of these two GNN paradigms from a technical perspective. Within each category of GNNs, there exists a multitude of theoretical underpinnings guiding their design. This diversity presents challenges in conducting a unified analysis across disparate GNN frameworks. To surmount this barrier, we will investigate a novel perspective we propose, which could potentially harmonize the treatment of both types within a singular framework, and subsequently discuss its merits and demerits.

• Current research of graph neural network (5 min)
• Overview of spatial-based GNN (5 min)
• Overview of spectral-based GNN (5 min)
• The challenge, benefit of studying the connection between spatial and spectral GNN (5 min)

2, Spectral graph theory, approximation theory and GNN (50 min)

In this section, we will elucidate the proposed unified perspective through case studies and comprehensive theoretical explanations. For instance, as a fundamental operation, graph matrix normalization will be examined from both spectral and spatial viewpoints, thoroughly justifying the necessity of normalization. Furthermore, we will delve into graph convolution from both perspectives: Spectral graph theory is committed to the analysis of graph structures via the eigenvectors and eigenvalues of matrices associated with the graph, notably the adjacency or Laplacian matrices. Approximation theory, which focuses on identifying the optimal approximation for a function within a designated class of functions, plays a pivotal role in revealing the expressiveness in terms of function of eigenvalues. By leveraging the spectral domain in conjunction with approximation methods, we can categorize all GNNs into linear, polynomial, and rational functions. Correspondingly, the spatial approach can be segregated into first-order, higher-order, and skip-connection strategies among neighbors. These two sets of three-tier categorizations precisely correspond, providing an additional layer of insights, rendering our proposed perspective both unified and conducive to elucidating a broader spectrum of GNNs.

• Spectral graph theory (15 min)
• Approximation theory for GNNs: linear, polynomial and rational functions (15 min)
• Case Study: normalization, GCN and DeepWalk (15 min)
• Unified framework to bridge spatial and spectral GNNs (5 min)

3, Theoretical study: viewpoints of uncertainty, sampling oversmoothing, and inverse (60 min)

This section offers a profound and innovative theoretical perspective on prevalent topics within the domain of Graph Neural Networks (GNNs), exploring areas such as uncertainty, sampling, oversmoothing, and inverse problems to furnish a comprehensive framework for analyzing and comprehending various phenomena in GNNs. Moreover, these new insights can be interpreted through both spectral and spatial lenses. The uncertainty principle is employed to understand the distinct global and local effects observable in the spectral and spatial domains, respectively. The examination of sampling theory in the context of graphs is facilitated by a unified framework that integrates explicit closed-form solutions. The phenomenon of oversmoothing in GNNs, characterized by the diminution of node feature distinctiveness as successive layers are applied, can be interpreted within the same conceptual framework. The realm of inverse problems, focusing on resolving edge-related challenges in graphs, presents an alternate viewpoint for understanding GNNs.

• Uncertainty principle: Global vs. Local Perspectives (10 min)
• Theoretical comparison between spatial and spectral methods (10 min)
• Sampling Point of View (15 min)
• Over-smoothing Point of View (15 min)
• Inverse Problem Point of View (10 min)

4, Future Directions (20 min)

Recent advancements in using Partial Differential Equations (PDEs) for graph analysis have shown promise. Specifically, the behavior of polynomial and rational functions in this context bears resemblance to diffusion and wave functions, offering innovative tools to study GNNs leveraging these functions. However, current spectral graph theories predominantly cater to simple graphs. Extending these theories to encompass signed, directed, and hypergraphs remains an underdeveloped area. While there has been progress, a comprehensive theoretical framework integrating these graph types is still in development. Future research endeavors will explore the feasibility of formulating a unified spectral theory. This theory would not only encompass various graph types, including simple, signed, directed, and hypergraphs, but also address the dynamics intrinsic to these graphs. The aim is to create a holistic and integrated theoretical understanding of graph behaviors and properties across a diverse range of graph types and their dynamics.

• PDE on graphs (10 min)
• Spectral Graph Theories for Non-simple graphs (5 min)
• Graph Dynamics (5 min)

5, Conclusion and Discussion (15 min)

• Conclusion (5 min)
• Q and A (10 min)

Content in this program

• Background

  Background of GNN.

• GNN

  About spectral graph theory.

• Theory

  Theoretic Study.

• Future Directions

  Discuss the future directions.

FAQs