Fig. 3
From: Designing a clustering algorithm for optimizing health station locations
Incorrectly detected clusters (+ sign signifies too many centroids,—sign too few) happen in k-means when there are many clusters and some of them are well separated. K-means may fail even with seemingly easy datasets (named A2, S2, Unbalance) to find all clusters correctly because the algorithm is incapable of moving the centroids across empty areas (deserts). Repeating the algorithm compensates for this only partly but relies too much on luck