Imba: Configuration-free Imbalanced Learning

Class imbalance of the target variable is a common feature of quite a few areas. Classic machine learning models are not the best solution in this case, since there will be a prediction bias towards the majority class. To solve this problem, various balancing techniques have been invented, grouped into: undersampling, oversampling and reweighting. However, their implementation requires manual research and configuration. In order to simplify the use of machine learning models, various methods and tools for automated machine learning are being developed. In this paper, the question of the applicability of existing methods to the problem of imbalanced classification was investigated. As it turned out, this problem is solved by them mainly by the same means as in a balanced classification setting. In this connection, Imba is announced - configuration-free imbalanced learning tool. The AutoML benchmark performance revealed worthy competition with the leading solution for automated machine learning model search and hyperparameter optimization - AutoGluon. But more importantly, these results were achieved with a search space of only three classifiers, resulting in significant reductions in computational costs and hence savings in operating time.