Double Cross Validation for Model Based Classification
Romain François, Florent Langrognet · 2006
where pk are the mixing proportions, μk ∈ R the mean vector of the k component, Σk its covariance matrix and Φ(.|μ,Σ) the normal probability density function with mean vector μ and variance matrix Σ. Celeux and Govaert (1995) proposed a decomposition of the variance matrices in terms of volume, orientation and shape. That decomposition yields 14 models from the simplest [λI] (same volume, shape and orientation with spherical variance matrices) to the standard QDA model [λkCk].