Class imbalance handling using wrapper-based random oversampling

We propose a novel algorithm for handling class imbalance. Class imbalance is a problem occurring in some valuable data such as medical diagnosis, fraud detection, oil spills, etc. The problem influences all supervised classification algorithms therefore a large amount of research is being done. The problem is tackled by preprocessing the data using wrapper-based random oversampling. Wrapper is a preprocessing approach that makes use of system (classifier) feedback to guide preprocessing. The wrapper approach is used to find regions suitable for sampling. Genetic algorithm is used as the basis of the wrapper approach to evolve the optimal regions. After specifying the optimal region random oversampling is performed to generate synthetic data. We evaluate our method using real world datasets with different imbalance ratios. We use two different classifiers that are Fisher and k-NN. The proposed algorithm is compared with two other oversampling methods namely SMOTE and random oversampling. The results show that the proposed algorithm is a suitable preprocessing method for handling class imbalance.