Cascade of phase transitions for multiscale clustering.

Area of Science:

• Data Science
• Machine Learning
• Statistical Modeling

Background:

• Clustering complex datasets with multiscale structures presents significant challenges.
• Existing methods often require prior knowledge of the number of clusters or scales.

Purpose of the Study:

• To develop a robust framework for clustering datasets exhibiting multiscale structures.
• To extract information about cluster number and size at different scales automatically.
• To learn principal graphs from spatially structured, multiscale data.

Main Methods:

• Utilizing a cascade of phase transitions during simulated annealing of the expectation-maximization algorithm.
• Employing weighted local covariance for feature extraction.
• Analyzing linear stability of the iterative scheme to identify phase transitions.
• Integrating regularized Gaussian mixture models.

Main Results:

• Successfully clustered datasets with inherent multiscale structures.
• Extracted a posteriori information on the number and size of clusters across different scales.
• Identified the threshold for the first phase transition and methods to approximate subsequent transitions.
• Learned principal graphs from spatially structured, multiscale datasets.

Conclusions:

• The proposed framework effectively handles multiscale data clustering without prior assumptions.
• The method provides valuable insights into the hierarchical structure of datasets.
• It offers a powerful approach for learning complex data representations, including principal graphs.