On a two-truths phenomenon in spectral graph clustering.

Area of Science:

• Graph theory
• Machine learning
• Neuroimaging

Background:

• Spectral graph clustering groups data based on graph structure without predefined labels.
• Commonly uses K-means or Gaussian mixture models with Laplacian spectral embedding (LSE) or adjacency spectral embedding (ASE).

Purpose of the Study:

• To investigate the
• two-truths
• phenomenon in spectral graph clustering, comparing LSE and ASE.
• To illustrate how different spectral embedding methods yield distinct clustering results.

Main Methods:

• Applied spectral graph clustering to a diffusion MRI connectome dataset.
• Utilized both Laplacian spectral embedding (LSE) and adjacency spectral embedding (ASE) for graph vertex clustering.
• Employed K-means clustering on the spectral embeddings.

Main Results:

• Demonstrated that LSE and ASE produce different clustering outcomes.
• LSE effectively captured left vs. right hemisphere structural affinities.
• ASE identified a core-periphery structure related to gray matter and white matter.

Conclusions:

• The choice of spectral embedding method significantly impacts clustering results in graph analysis.
• LSE and ASE capture complementary structural information in diffusion MRI connectomes.
• Understanding these differences is crucial for accurate data interpretation in neuroimaging and other fields.