Transfer Learning for Classification on Unbalanced Data

Based on TrAdaboost, we propose an Unbalanced TrAdaboost (UBTA) algorithm which is a binary classification algorithm aiming to classify unbalanced data. UBTA calculates the weights of weak classifiers using the auprc (the Area Under the Precision-Recall Curve) of different classes and updates the weights of misclassified samples of different classes with different schemes. In combination with G-mean and BER, the AUC measure is more accurate when evaluating the performance of unbalanced classification since it is insensitive to changes in class distributions. The experiments indicate that the UBTA algorithm achieves better results for unbalanced data sets, strengthens the attention to the minority instances and maintains high classification accuracy for majority instances.