Genetic programming for development of cost-sensitive classifiers for binary high-dimensional unbalanced classification

Abstract

Abstract Genetic programming (GP) has the built-in ability for feature selection when developing classifiers for classification with high-dimensional data. However, due to the problem of class imbalance, the developed classifiers by GP are prone to be biased towards the majority class. Cost-sensitive learning has shown to be effective in addressing the problem of class imbalance. In cost-sensitive learning, cost matrices are often manually designed and then considered by classification algorithms to treat different mistakes differently. However, in many real-world applications, cost matrices are unknown because of the limited domain knowledge in complex situations. Therefore, in this paper, we propose a novel GP method to develop cost-sensitive classifiers, where a cost matrix is automatically learned, instead of requiring it from domain experts. The proposed method is examined and compared with existing methods on ten high-dimensional unbalanced datasets. Experimental results show that the proposed method outperforms the compared GP methods in most cases.