[Paper Review] Understanding the Basis of Graph Convolutional Neural Networks via an Intuitive Matched Filtering Approach

Ljubiša Stanković, Danilo P. Mandic|arXiv (Cornell University)|Aug 23, 2021

Advanced Graph Neural Networks20 references4 citations

TL;DR

This paper reframes Graph Convolutional Neural Networks (GCNNs) through the lens of matched filtering, demonstrating that GCNN convolution layers detect features in graph-structured data by correlating inputs with learned templates—akin to matched filters. The approach unifies GCNN operations, including ReLU and max-pooling, under this framework and shows that standard CNNs and fully connected networks are special cases, with a numerical example illustrating signal propagation and learning step-by-step.

ABSTRACT

Graph Convolutional Neural Networks (GCNN) are becoming a preferred model for data processing on irregular domains, yet their analysis and principles of operation are rarely examined due to the black box nature of NNs. To this end, we revisit the operation of GCNNs and show that their convolution layers effectively perform matched filtering of input data with the chosen patterns (features). This allows us to provide a unifying account of GCNNs through a matched filter perspective, whereby the nonlinear ReLU and max-pooling layers are also discussed within the matched filtering framework. This is followed by a step-by-step guide on information propagation and learning in GCNNs. It is also shown that standard CNNs and fully connected NNs can be obtained as a special case of GCNNs. A carefully chosen numerical example guides the reader through the various steps of GCNN operation and learning both visually and numerically.

Motivation & Objective

• To demystify the inner workings of GCNNs, which are often treated as black boxes despite their growing use in irregular domain data processing.
• To establish a theoretical foundation for GCNNs by linking their convolution operation to matched filter theory, a well-understood concept in signal processing.
• To unify the understanding of GCNN components—such as ReLU activation and max-pooling—within the matched filtering framework.
• To demonstrate that standard CNNs and fully connected networks are special cases of GCNNs under specific graph structures.
• To provide a pedagogical, step-by-step numerical example that visually and numerically illustrates GCNN operation and information propagation.

Proposed method

• Reinterprets GCNN convolution as matched filtering by showing that the convolution operation with a time-reversed kernel is equivalent to cross-correlation, the core of matched filtering.
• Uses the matched filter output formula $ y(n) = x(n) * w(-n) $ to formalize feature detection in graph signals, where $ w(n) $ is the feature template.
• Applies the matched filter framework to explain nonlinear ReLU and max-pooling operations as part of the feature detection and selection process.
• Introduces indicator matrices $ \mathbf{P} $ and $ \mathbf{M}^{ReLU+MP} $ to track which signal components survive nonlinear and pooling operations, enabling gradient backpropagation tracking.
• Employs graph coarsening and lifting via pseudo-inverse $ \mathbf{P}^+ $ to model hierarchical feature learning across multiple scales, with $ \mathbf{W}_c = \mathbf{P} \mathbf{W} \mathbf{P}^T $ and $ \mathbf{W}_L = \mathbf{P}^+ \mathbf{W}_c (\mathbf{P}^+)^T $.
• Demonstrates that standard CNNs and fully connected networks emerge as GCNNs when the graph is fully connected or has a specific structure, respectively.

Experimental results

Research questions

• RQ1 How can the operation of GCNNs be interpreted through the established principles of matched filtering in signal processing?
• RQ2 To what extent can the nonlinear activation (ReLU) and pooling (max-pooling) layers in GCNNs be explained as part of a matched filtering process?
• RQ3 Can standard convolutional and fully connected neural networks be derived as special cases of GCNNs under specific graph topologies?
• RQ4 How does information propagate through a GCNN, and how can this be visualized and tracked using indicator matrices?
• RQ5 What is the role of graph coarsening and lifting in hierarchical feature learning within the matched filtering framework?

Key findings

• GCNN convolution layers perform matched filtering by correlating input signals with learned feature templates, with the kernel acting as a time-reversed template for optimal detection.
• The ReLU and max-pooling operations are naturally explained within the matched filtering framework as mechanisms that preserve only the most relevant feature responses.
• The indicator matrix $ \mathbf{M}^{ReLU+MP} $ successfully identifies which signal components survive nonlinear and pooling operations, enabling accurate gradient backpropagation.
• Graph coarsening and lifting via $ \mathbf{P}^+ $ allow hierarchical feature learning, with the weight matrix transformation $ \mathbf{W}_c = \mathbf{P} \mathbf{W} \mathbf{P}^T $ preserving structural relationships across scales.
• Standard CNNs and fully connected networks are shown to be special cases of GCNNs when the underlying graph is fully connected or has a specific node arrangement.
• The numerical example in the paper demonstrates that GCNNs detect localized features across irregular graphs using matched filtering principles, with clear visual and numerical tracking of signal propagation and learning.