Cascade evaluation

In fact, what we want to evaluate is how well a given clustering algorithm is able to capture interesting, meaningful and usable information. Meaningful information in that case correspond to a new knowledge that is interesting to use for some purpose. We also expect the algorithm to be able to capture such interesting information on various types of datasets. Based on these considerations, we propose a new methodolgy for clustering evaluation and clustering comparison. The main idea is to use the clustering results to enrich a given dataset, and check if this extra-knowledge is somewhat useful for the apprehension of the dataset. In particular, we conjecture that, if the results of a given supervised learning algorithm are improved when new information are added to a dataset, from the results of a clustering process, then it means that such information were somewhat useful. And thus, the clustering results can be regarded as relevant. Figure 1 summarizes the main steps of our proposed methodology. An example of information that can be added from clustering results is the membership of the data points to the clusters, if the output of the clustering is a partition of the dataset. Another possible way to add information from clustering results is to associate to each data point a set of attributes representing the center of the cluster it belongs to. This research is inspired by the works on classifier combinations, and in particular cascade generaliza-