Imbalanced Data Learning: Importance, Recent Trends, and Challenges

Abstract

Imbalanced data learning resides among the top research area for many decades and it is still very important in real-world domains like medical diagnosis, chemical reaction analysis, the judicial system, road accident analysis, etc. Even after a decade of work, the conventional algorithms show poor accuracy, particularly for highly imbalanced data. Existing techniques of imbalanced data learning are mostly based on sampling methods. High imbalanced ratio, class overlapping makes imbalanced learning more challenging. We have studied imbalanced data learning, recent work on imbalanced data, and discuss the challenges that we face while learning from imbalanced data. We experiment with conventional algorithms on balanced and unbalanced data and compare the results. We found that conventional algorithms show poor performance for the unbalanced data.