Pruning Optimum-Path Forest Classifiers Using Multi-Objective Optimization

Multi-objective optimization plays an important role when one has fitness functions that are somehow conflicting with each other. Also, parameter-dependent machine learning techniques can benefit from such optimization tools. In this paper, we propose a multi-objective-based strategy approach to build compact though representative training sets for Optimum-Path Forest (OPF) learning purposes. Although OPF pruning can provide such a nice representation, it comes with the price of being parameter-dependent. The proposed approach cope with that problem by avoiding the classifier to be hand-tuned by modeling the task of parameter learning as a multi-objective-oriented optimization problem, which can be less prone to errors. Experiments on public datasets show the robustness of the proposed approach, which is now parameterless and user-friendly.