Selection of the Best Base Classifier in One-Versus-One Using Data Complexity Measures.

When dealing with multiclass problems, the most used approach is the one based on multiple binary classifiers. This approach consists of employing class binarization techniques which transforms the multiclass problem into a series of binary problems which are solved individually. Then, the resultant predictions are combined to obtain a final solution. A question arises: should the same classification algorithm be used on all binary subproblems? Or should each subproblem be tuned independently? This paper proposes a method to select a different classifier in each binary subproblem-following the one-versus-one strategybased on the analysis of the theoretical complexity of each subproblem. The experimental results on 12 real world datasets corroborate the adequacy of the proposal when the subproblems have different structure.