Ensemble Genetic Programming for Classifying Gene Expression Data

Ensemble is a representative technique for improving classification performance by combining a set of classifiers. It is required to maintain the diversity among base classifiers for effective ensemble. Conventional ensemble approaches construct various classifiers by estimating the similarity on the output patterns of them, and combine them with several fusion methods. Since they measure the similarity indirectly, it is restricted to evaluate the precise diversity among base classifiers. In this paper, we propose an ensemble method that estimates the similarity between classification rules by matching in representation-level. A set of comprehensive and precise rules is obtained by genetic programming. After evaluating the diversity, a fusion method makes the final decision with a subset of diverse classification rules. The proposed method is applied to cancer classificaiton using gene expression profiles, which requires high accuracy and reliability. Especially, the experiments on popular cancer datasets have demonstrated the usefulness of the proposed method.