Gaussian mixture clustering — a probability per cluster
A Gaussian mixture fits k ellipsoidal densities and assigns each row a responsibility, the probability it was drawn from each density. The label you see is the density with the highest responsibility. Overlapping groups stay overlapping: a row can be 0.6 / 0.4 instead of a hard 1 / 0. Full covariance lets each cluster tilt, so the ellipse need not line up with the axes.
On scaled Iris, 3 full-covariance components and seed 42 reached adjusted Rand 0.904 against species, the best species recovery of the Iris runs, at a silhouette of only 0.374. Silhouette penalizes the overlap that the mixture is willing to model. On wine the same setup reached adjusted Rand 0.929. If you only rank models by silhouette, you will throw away the fit that matched the held-out groups.
Full covariance estimates a matrix per component. With few rows and many columns that matrix becomes singular; ridge-like regularization inside the fit, or a tied covariance, is what keeps it solvable.
Settings that change this fit
- n_components
- How many Gaussians are mixed. This is k, but each component has a mean, a covariance, and a weight.
- covariance
- Full allows tilted ellipses. Diagonal forces axis-aligned ellipses. Spherical forces balls, which is close to K-Means.