A Study of k-means Method where Starting Conditions are Changed: A Simulation Study Rok Podgornik, Marijan Zafred, and Anja Pajtler Abstract The goal of our work is to evaluate the performance of k-means clustering method. We created three (or two) groups of units, which compose one of our three different structures. First we created groups, which were clearly separated in three-dimensional space, and then we started to decrease the differences among them with increasing the data dispersion at keeping the centroids constant. We wanted to find out how long we can continue with this process until we still get satisfactory results at clustering of units into groups. Six different methods of clustering were used (with different initial cluster centers) and we evaluated every one of them. In comparison with initial clustering we are observing three criteria: percentage of correctly clustered units, centroids dislocation and value of criteria function. The results of clustering are presented graphically with added values of three chosen criteria for better possibility of evaluation. For the simulation of the data and graphical presentation we used program package R. |