Diversity in classifier ensembles: fertile concept or dead end?

Diversity is deemed a crucial concept in the field of multiple classifier systems, although no exact definition has been found so far. Existing diversity measures exhibit some issues, both from the theoretical viewpoint, and from the practical viewpoint of ensemble construction. We propose to address some of these issues through the derivation of decompositions of classification error, analogue to the well-known bias-variance-covariance and ambiguity decompositions of regression error. We then discuss whether the resulting decompositions can provide a more clear definition of diversity, and whether they can be exploited more effectively for the practical purpose of ensemble constructio