Classifier Ensemble Selection And It's Application To Analyze Gene Data

Ensemble learning method uses suitable methods to integrate the results of different classifiers of the same problem. Compared with single classifier, it will achieve better accuracy and robustness. But the classifiers in ensemble classifier are not all positive impact on the result. Ensemble selection is try to select the best classifier subset, in order to improve the ensemble classifier performance and reduce memory requirements and computing cost.Two ensemble selection methods are proposed in this paper.One is static ensemble selection method based on kappa coefficient, the other is dynamic ensemble selection method based on firefly algorithm. These two methods were applied in different situations. Static ensemble selection method is suitable for small data, and dynamic ensemble selection method is suitable for big data. Before classifier selection is applied, classification-related genes are chosen using a ranking aggregation technique, then these genes are divided into groups by an affinity propagation clustering algorithm. Some diversity and distinguishing gene subsets are constructed by randomly selecting a gene from each group and are used to train base classifiers. After obtained some base classifiers, the first method is selecting classifier which kappa coefficient greater than kappa threshold, the second method is using similar clustering method to select classifiers with higher accuracy and the difference between classifiers is larger.The experimental results on five gene data show that the two methods obtained better accuracy than classical approach. The results of the two methods in this paper were almost the same. The first method can quickly select the appropriate category subset, the second method can save many time when the data is big and obtain better classification results.