A Hybrid Resampling Approach to Handle Class Imbalance Problem and Missing Data
Pranita Baro, Malaya Dutta Borah · 2022 IEEE 9th Uttar Pradesh Section International Conference on Electrical, Electronics and Computer Engineering (UPCON) · 2022
Class imbalance and missing data is a frequent issue in machine learning, particularly in classification issues. When working with real-life datasets to train or test classification methods, there is a possibility to run into class imbalance and incompleteness of data. Both scenarios can frequently be found in the same dataset. Class imbalance, as well as missing data, imposes limitations on the ability of classifiers to evaluate and predict the correct class. This work provides a hybrid combination of oversampling and undersampling strategies for dealing with class imbalance issues and data incompleteness. First, an oscillator built using a factor is presented that quantifies how much missing data belongs to the majority class compared to those on the minority class. The training dataset is then divided into several subsets, and the oscillator, whose value varies between 0 and 100, is used to determine whether to apply undersampling or multiple imputations based on oversampling to each subset. The effectiveness of the proposed method has been validated after experiments on benchmarked datasets.