Efficient Greedy Unsupervised Learning of Multivariate Mixture Models with Moment-Based Splitting

Unsupervised clustering using finite mixture models has been widely applied over the past decades. Gaussian mixture densities allow a great generalization of the input data. However, their computational complexity increases with the number of mixture components. Greedy algorithms address this issue by iteratively splitting components of the mixture, but they heavily depend on the splitting method, which often leads to suboptimal solutions. We propose a novel splitting rule that exploits the third and fourth statistical moments - skewness and kurtosis - iteratively refines Gaussian components. This method ensures computational efficiency, robust parameter tuning, and satisfactory accuracy. Experimental results demonstrate the speed and the reliability of this approach, making it suitable for large-scale applications.