An ensemble-clustering-based distance metric and its applications

A distance metric learned from data reflects the actual similarity between objects better than the geometric distance. So, in this paper, we propose a new distance that is based on clustering. Because objects belonging to the same cluster usually share some common traits even though their geometric distance might be large. Thus, we perform several clustering runs to yield an ensemble of clustering results. The distance is defined by how many times the objects were not clustered together. To evaluate the ability of this new distance to reflect object similarity, we apply it to two types of data mining algorithms, classification (kNN) and selective sampling (LSS). We experimented on standard numerical datasets and on real colour images. Using our distance, the algorithms run on equivalence classes instead of single objects, yielding a considerable speedup. We compared the kNN-EC classifier and LSS-EC algorithm to the original kNN and LSS algorithms.