Comparison of tree assignments, k -means and horizontal cut. Clustering the training dataset into 14 classes using k-means (left) or Ward’s clustering using a horizontal cut (middle) leads to partitionings with a Rand index of around.9 with relatively high standard deviation. A partitioning obtained from Ward’s clustering using tree-assignments leads to a significantly higher Rand index (right). Note that the Rand index approaches 1 for datasets with many classes. Yet, the difference after Monte-Carlo type validation is clearly significant.