5. Discussion
This paper presented a novel version of the K-means algorithm based on the Mahalanobis distance metric. While the
use of Mahalanobis distances is not new in clustering framework, they are not commonly used due to the necessity to
initialize data group covariance matrices. We proposed a strategy aiming at addressing this issue. The developed procedure
is illustrated on a synthetic dataset; its performance was compared to that of two other versions of the K-means algorithm.
A conducted simulation study proved viability of the proposed method, which outperformed the competitors in multiple
situations. An application of the procedure to the time-honored Iris dataset was considered with good results.