Hybrid species tend to exhibit a mixture of parent characteristics; we propose chimeral clusters as exhibiting a mixture of parent parameters, a type of intercluster structure. Morphometric measurements in the iris dataset describe the hybrid Iris versicolor as intermediate to those of parent species Iris setosa and Iris virginica, which motivates our extension of Gaussian mixture models to allow mixing in the parameter space. We propose a mixing mechanism whereby chimeral clusters are parameterized by a convex combination of fully varying prototype cluster parameters and characterize the identifiability of the postulated mixture model. Estimation of chimeral clustering models is described using variations of the expectation-maximization algorithm and the solution to the continuous-time algebraic Riccati equation. The efficacy of chimeral clustering is demonstrated using morphometric datasets describing iris, Cooper's hawks, and water striders, with comparisons to typical Gaussian mixture models. We evaluate parameter recovery on a synthetic dataset and demonstrate that parsimonious covariance matrices and chimeral clustering capture different kinds of intercluster structure.
展开▼
