Coarse-to-fine classifier ensemble selection using clustering and genetic algorithms

A good classifier ensemble should show high complementarity among classifiers to produce a high recognition rate and it should also have a small size to be efficient. This paper proposes a classifier ensemble selection algorithm operating in a coarse-to-fine paradigm. For the algorithm to be successful, the original classifier pool should be sufficiently diverse. So this paper produces a large classifier pool by combining several different classification algorithms and several feature subsets. The coarse selection stage reduces greatly the size of the classifier pool using a clustering algorithm. The fine selection finds the near-optimal ensemble using genetic algorithms. A hybrid genetic algorithm with improved searching capability is also proposed. The experimentation used handwritten numeral datasets and UCI datasets. The experimental results and the test of statistical significance showed that the proposed algorithm is superior to the conventional ones.