A Dual Sampling Approach for Improved Classifier Performance on Imbalance Datasets

Background: The inability of traditional machine learning models to adequately classify minority instances in imbalanced datasets is a known challenge that militate against the successful application of these models in several real-world domains. To address this problem, several techniques including data sampling are mostly used. Though reducing the imbalance ratio via sampling is reported to improve classifier performance, most approaches do not consider the intra-class distribution of instances while sampling, which often lead to loss of significant information or on the contrary cause data redundancy. Methods: This study proposes a novel Dual Sampling Technique (DST) that minimises these challenges and enhances classifier performance on imbalance datasets. The technique proceeds by first clustering a training set into a number of clusters determined a priori using the elbow method. Sampling ratios are computed from each cluster and either random undersampling or a novel average oversampling technique or both are used to perform sampling in each cluster depending on the imbalance ratio. The resulting datasets are used to train Random Forest, Decision Tree and K-Nearest Neighbor classifiers and their performance evaluated. Findings: Experimental results showed that the performance of the classifiers significantly improved in most cases when the proposed technique is used to sample the training set prior to model building than when Random Undersampling (RUS), Random Oversampling (ROS), Synthetic Minority Oversampling Technique (SMOTE) and Cluster-Based Undersampling (CBU) are used. Novelty: The novelty of the proposed technique lies in the exploration of a unique concept that sought to minimise the imbalance ratio in datasets while maintaining their natural distribution by uniquely performing both undersampling and oversampling on the same dataset.