This approach works well if the number of clusters is not too large and they are all fairly well represented. But even in the case of n=1500 points spread among K=15 clusters in equal proportions, the probability of having a “seed” in each cluster is approximately 3.2⋅10−6. The need for the initialization techniques that would choose initial cluster centers in a more careful way rather than randomly is addressed, for instance, by Erisoglu et al. (2011). Proposed variations of the algorithm typically aim at finding better original cluster centers but do not concern the shape of detected clusters. In the mean time, the ability of a particular distance measure to accommodate clusters of certain configuration is another important aspect arising in the application of the K-means algorithm. The most widely used, Euclidean distance measure, works well for clusters with roughly spherical homogeneous covariance matrices. The use of the Mahalanobis distance