Beyond pairwise dependence: A multi-filter fusion network for graph representation learning.

Area of Science:

• Graph Neural Networks
• Spectral Graph Theory
• Machine Learning

Background:

• Spectral Graph Neural Networks (GNNs) commonly use polynomial filters based on pairwise network structures.
• These methods face limitations such as over-smoothing, over-squashing, and overfitting due to reliance on first-order graph properties.
• Existing approaches struggle to capture complex, high-order relationships within graph data.

Purpose of the Study:

• To address the limitations of traditional spectral GNNs.
• To propose a novel network architecture capable of learning arbitrary graph spectrum filters.
• To enhance the ability of GNNs to model complex graph structures and signals.

Main Methods:

• Introduced the Multi-Filter Fusion Network (MFFN) architecture.
• Utilized Fourier expansion on diverse Laplacian matrices of order-incidence.
• Encoded eigenvalues with simplicial complexes to capture higher-order graph interactions.
• Employed semantic-adaptive masks to approximate multiple spectral filters.

Main Results:

• MFFN demonstrated the capability to learn arbitrary graph spectrum filters.
• Achieved superior performance in node-level classification tasks.
• Achieved superior performance in graph-level classification tasks.
• Effectively captured interactions beyond pairwise dependence using simplicial complexes.

Conclusions:

• The proposed MFFN effectively overcomes the limitations of conventional spectral GNNs.
• MFFN offers a powerful approach for modeling complex graph signals and learning higher-order graph structures.
• The method shows significant promise for advancing graph representation learning in various applications.