Turtle shell clustering: A mixture approach to discriminative clustering with applications to flow cytometry and other data

Researchers Mackenzie R. Neal, Paul D. McNicholas, and Arthur White have introduced 'turtle shell clustering,' a fully unsupervised probabilistic method that blends generative and discriminative clustering approaches. Traditional generative methods describe cluster shapes (e.g., Gaussian distributions), while discriminative methods define boundaries between clusters. Turtle shell clustering combines both by using a regularized mutual information objective function, with a mixture of mixtures of Gaussian and uniform distributions for the conditional model. This allows it to automatically select the number of clusters via a regularizing term and a merge step, inspired by reversible jump Markov chain Monte Carlo methods in Bayesian clustering. The method can estimate non-linear boundary lines, making it robust to noise and irregular cluster shapes.

The researchers tested turtle shell clustering on simulated and real datasets, including flow cytometry data—a common challenge in biology where data can be noisy and clusters irregular. Results showed it outperforms traditional methods (e.g., k-means, Gaussian mixture models) in capturing intuitive clusters, especially with outliers or non-spherical shapes. This work, published on arXiv (2604.23083), has implications for fields like bioinformatics, medical diagnostics, and any domain requiring robust unsupervised pattern recognition. By automating cluster number selection and handling complex data structures, turtle shell clustering could reduce manual tuning and improve insights from high-dimensional datasets.