Matched signal detection on graphs: theory and application to brain network classification.

Area of Science:

• Graph Signal Processing
• Statistical Signal Detection
• Machine Learning

Background:

• Traditional signal detection methods often overlook intrinsic data structures.
• Graph-based data representations are increasingly important in various scientific fields.

Purpose of the Study:

• To develop a matched signal detection (MSD) theory for signals with intrinsic graph structures.
• To extend existing detection frameworks to accommodate graph-based signal properties.
• To evaluate the performance of the proposed MSD theory on real-world data, specifically for network classification in Alzheimer's disease (AD).

Main Methods:

• Formulation of hypothesis tests for deterministic signals in subspaces spanned by graph Laplacian eigenvectors.
• Development of a weighted energy detector for smooth signals with negligible noise variance.
• Extension to signals with prior distributions using an Ising model, leading to a test statistic based on signal variations on graph structures.
• Evaluation using simulations and real-world network data for Alzheimer's disease classification.

Main Results:

• The proposed MSD theory effectively handles graph-structured signals.
• The method demonstrates superior performance compared to traditional Principal Component Analysis (PCA) in network classification tasks.
• Preliminary results show successful exploitation of the sub-manifold structure of data.

Conclusions:

• The developed matched signal detection theory provides a powerful framework for analyzing signals with inherent graph structures.
• This approach offers significant advantages over conventional methods, particularly in complex network analysis and disease classification.
• The MSD theory's ability to leverage data's sub-manifold structure opens new avenues for advanced signal processing and machine learning applications.