Gaussian Mixture Models

Good example


Good notebook

[] []

GMMs assume all data points are mixture of Gaussian distributions, with unknown parameters.

It’s actually quite similar to K-means, but unlike K-means, GMMs can learn clusters with any elliptical shape, not just circles. Also, GMMs give probabilities of belonging to a cluster, not hard assignment. This means a point can belong 50% to one cluster and 40% to another.

GMM can also be useful for outlier detection. Points with low likelihoood can be labeled as outliers.