Hodnocení úspěšnosti metod při shlukové analýze

The cluster method is one of the multivariate methods. Its goal is to classify objects into groups (clusters) according to their similarity in one cluster and the difference from other clusters. This diploma thesis where the analysis is performed on the basis of 25 files from the Internet database "UCI Machine Learning Repository" compares the success of clustering objects using the original variables including the correlated and created main components. The first variant classifies objects with 68.97 % of success on the average. In contrast, the success rate of the classifying objects using components created by the main component method is 65.99 %. This work also compares the success of classifying objects into the clusters, between five specific methods of the cluster analysis. In this context, it is found that in both cases of analysis, the most successful method is the Ward´s method. Its average success rate for classifying objects based on the original variables is 85.33 % and based on the variables obtained using the principal components method, it is 5.17 percentage points lower. When comparing the success of methods evaluated on the basis of the original variables, it is found that the farthest neighbor method correctly classifies 72.87 % of the objects. Other methods with lower success are the median method (63.76 %) and the centroid method (63.75 %). The least successful method is the nearest neighbor method (59.15 %). When clustering objects using components obtained from principal components method, it is found that the order of success of the methods it is the same as in the previous case. The farthest neighbor method shows 66.38% success rate of classifying objects into the clusters. The less successful method are the median method (62.95 %) and the centroid method (61.43 %). The lowest success rate is again shown by the nearest neighbor method (59 %).