Multiobjective genetic classifier selection for random oracles fuzzy rule-based classifier ensembles: How beneficial is the additional diversity?

Recently we proposed the use of the Random Linear Oracles classical classifier ensemble (CE) design methodology in a fuzzy environment. It derived fuzzy rule-based CEs obtaining an outstanding performance. Random Oracles introduce an additional diversity into the base classifiers improving the accuracy of the entire CE. Meanwhile, the overproduce-and-choose strategy leads to a good accuracy-complexity trade-off. It is based on the generation of a large number of component classifiers and a subsequent selection of the best cooperating subset of them. The current contribution has a twofold aim: (1) Introduce a new Random Oracles approach into the fuzzy rule-based CEs design; (2) Incorporate an evolutionary multi-objective overproduce-and-choose strategy to our approach analyzing the influence of this additional diversity in the final CE performance (focusing on the accuracy). To do so, firstly, we incorporate the two Random Oracle variants into the fuzzy rule-based CE framework. Then, we use NSGA-II to provide a specific component classifier selection driven by three different criteria. Exhaustive experiments are carried out over 29 UCI and KEEL datasets with high complexity (considering both the number of attributes as well as the number of examples) showing the good performance of the proposed approach.