A bi-objective hybrid algorithm for the classification of imbalanced noisy and borderline data sets

Abstract

Classification of imbalanced data sets is one of the significant problems of machine learning and data mining. Traditional classifiers usually produced suboptimal results for imbalanced data sets. This study proposed an idea of using a newly proposed bi-objective hybrid algorithm for the given classification task of binary imbalanced noisy and borderline data sets. The bi-objective hybrid algorithm was based on the hybridization of two metaheuristics, namely cuckoo search and covariance matrix adaptation evolution strategy. The validation of this proposed hybrid algorithm was confirmed in terms of the Pareto fronts. Thereafter, this algorithm was used in a methodology proposed for the classification task of the binary imbalanced data sets. The proposed methodology was based on an idea of estimating the probabilities from both classes (majority and minority) of a data set, using normal distribution. Optimization of parameters of the normal distribution was done with the help of the proposed algorithm. Different data sets (simulated, noisy borderline and real) were used. Four well-known classifiers with a preprocessing algorithm were cast-off for the comparison purpose. Performances of all classifiers were evaluated using three evaluation measures, sensitivity, G mean and F measure. A promising performance of proposed methodology was observed.