A Modified Version of K-Means Algorithm.

In this work is presented a modified version of the K-Means which identifies cluster stability. The stability is defined by a threshold based on a percentage of the largest displacement of centroid at first iteration. A cluster is considered stable when the largest centroid displacement in the current iteration achieves the 10% of threshold, and objects that remains in the same cluster in two consecutive iterations are removed from the classification phase in subsequent iterations. Eight different instances were used to validate the proposal, three synthetics and five reals. The modified version was compared against the standard and three related work versions. Results shows that the proposal reduced the execution time up to 92.14% regarding the standard version with only a 3.73% in the quality reduction. Despite the new version do not has the major reduction time in all cases, the algorithm reaches the best values for quality of grouping.