Class imbalance in the technology of concrete historical investigation support

The article continues a series of works devoted to the technology of concrete historical investigation support, built on the principles of co-creation and crowdsourcing and designed for a wide range of nonprofessional historians and biographers users. The article is devoted to the further development of the topic of data preparation for machine learning algorithms used in the technology. The special importance of binary classification for concrete historical research is shown. The problem of class imbalance in binary classification using machine learning algorithms and its consequences are described. It is shown that concrete historical data can be highly imbalanced. An overview of approaches to solving the problem of class imbalance elimination is given. The analysis of the specifics of concrete historical data was carried out, and on its basis, the oversampling approach was chosen as the most suitable for the technology. Algorithms implementing this approach are described; their advantages and disadvantages are evaluated. The ADASYN algorithm has been selected as the most promising for use in the technology conditions. The possibilities of the already included in the technology means of data noise and outliers control to compensate such a disadvantage of the ADASYN algorithm as sensitivity to outliers are evaluated.