IMPROVING RESULTS IN MULTI-CLASS CLASSIFICATION FOR IMBALANCED DATA BY ASSIGNING WEIGHTS TO CLASSES BASED ON MACHINE LEARNING MODELS

This study analyzed strategies for assigning weights to classes in the problem of imbalanced classification and elucidated the theoretical and practical aspects of applying these strategies in ensemble and neural network models. The imbalance in the proportion of classes for rare positive cases of medical diagnostics, the low result of the accuracy indicator, as a result, leads to a decrease in the results of the assessment criteria of the minority class recall in the models and is eliminated by assigning weight to the classes of the models to prevent the occurrence of an incorrect diagnosis. This study describes the weighted gradient-Hessian approach for gradient boosting family algorithms such as LightGBM and HistGradientBoosting, as well as weighted cross-entropy and, when necessary, threshold-moving and calibration methods for MLP. In mathematical form, the updating of leaf values using the Newton step, the calculation of cumulative gradients at the histogram bin level, and cross-entropy weighting for MLP based on output logits are presented. Additionally, boosting and weighted derivatives are included.