A Multi-dimensional Survey on Learning from Imbalanced Data

The problem of data imbalance in machine learning is related to the uneven distribution of the training examples to the involved classes. Nowadays, a large number of research fields and applications suffer from class imbalance, including Cybersecurity, Bioinformatics, Natural Language Processing, management of multimedia content, and so on. Directly training machine/deep learning classifiers with such data has been proved quite problematic, because the generated models become strongly biased towards the majority class. Unable to learn the minority classes effectively, the accuracy of these “imbalanced” models degrades rapidly. Given the importance of the problem, numerous researchers have introduced innovative, state-of-the-art approaches with the aim of addressing it. In this chapter, we present a survey on the most significant advances in the area, by adopting a multi-dimensional categorization approach. Specifically, we classify the relevant works according to: i) the application field they focus on, ii) the methods they introduce to mitigate class imbalance, and iii) the classification models they utilize to evaluate the introduced algorithms. Additionally, we cover the state-of-the-art overviews in a systematic manner and we describe the proposed methods and their achieved results.