Detailed Performance Study of Data Balancing Techniques for Skew Dataset Classification

Many real-world classification problems involve changing events where one class has comparatively fewer samples called minority class which is more important to detect. Consequently, the dataset is often unbalanced and shows significantly skewed data. Since the majority class dominates the learning process and tends to sketch all predictions, the conventional classification model leads to biased results where it may easily display excellent performance in the dominant class and bad performance in the minority class. Additionally, the traditional accuracy score is inaccurate since it assigns equal weight to actual positives and actual negatives. This study is aimed to present an empirical analysis of the data imbalance effect on classification algorithms. Six popular and effective data balancing techniques are applied to eight benchmark skewed datasets from the KEEL and Kaggle repositories to analyze and compare the performance.