A Two Tier Iterative Ensemble Method To Tackle Imbalance In Multiclass Classification

In real world applications, it is very common that the data skewness occurs among multiple classes. Several studies and various attempts were made in the past to overcome this imbalance problem which is a serious issue to the standard machine learning techniques especially classification and regression but, still there exists a need to handle the imbalance problem effectively. Datasets which are imbalanced generally include safe and unsafe minority samples. Our proposed approach is a classifier independent two tier iterative ensemble approach which focuses the rare minority sample's influence on learning from imbalanced datasets. Most of the informed oversampling techniques like SMOTE and its variants cannot be applied directly on rare class samples especially when the count of rare samples is too low. To alleviate this problem, in our proposed approach to learn from rare and outlying samples we proposed a hybrid oversampling technique used at different levels and make them balanced. The goal is to tone down the data imbalance at the data preprocessing stage itself by correcting or balancing the training data sets before moving to the learning part which makes the classifier to focus on its primary role and thereby it improves the learning process. The proposed two tier iterative ensemble approach shows a much significant improvement in the learning process among the multiclass imbalanced data which is clearly evident with the experimental results.