Harmonization Centered Ensemble For Small And Highly Imbalanced Medical Data Classification

Jiaru Li, Bowen Xin, Ziyi Yang, Junyan Xu, Shaoli Song, Xiuying Wang

Abstract: Deficient and highly imbalanced medical data often makes it challenging for accurate classification without bias. In this paper, we propose an ensemble learning framework to harmonize the discriminating information between the majority class and minority class, which is called Harmonization Centered Ensemble (HCE). The proposed framework leverages the classification hardness and neighborhood information concept during the under-sampling procedure to emphasize the importance of the minority class. It also applies the boosting technique, which refines the ratio of easily distinguishable and difficult to distinguish the majority class to calibrate the deficient attention on majority class. Extensive experimental results indicate that our framework outperforms state-of-the-art under-sampling ensemble frameworks on two highly imbalanced medical datasets with small data size, and offers more balanced performance for different target classes.