Strategic Blended Learning for Addressing Multiclass Imbalance in Classification

Numerous practical applications involve multiclass classification challenges, where the spread of data across the classes is frequently uneven. Due to this inherent data imbalance, supervised learning models often exhibit a bias toward classifying instances into the category with the highest number of data points, posing a significant problem that requires attention. In the realm of classification of data with multiple classes and imbalanced distributions, researchers in machine learning persistently endeavor to mitigate the classifier's inclination towards dominant class with more sample size. Existing studies aim to alleviate this inclination by alternatively adjusting the data for balance prior to the classifier's learning process or adjusting the classifier's training process to focus more on rare class having least number of samples, or employing hybrid strategies. Existing learning approaches encounter challenges in delineating clear boundaries between samples of classes resulting from an imbalanced class distribution and issues with class overlapping. Consequently, the recognition rate for minority classes is hindered. A novel Strategic Blended Learning (SBL) approach is introduced, employing dual decision trees. The first decision tree is utilized to generate an induced dataset through a grouping approach, while the second decision tree is employed to train and make decisions guided by the induced dataset. A noteworthy aspect of this algorithmic approach is its ability to identify rare data points without affecting the inherent distribution of data. To evaluate the proposed algorithm, various multiclass imbalanced datasets from UCI were employed for classification. The findings revealed that the Strategic Blended Learning method excels in detecting both rare and dominant category samples.