Розробка економічно чутливої оптимізації штучних нейронних мереж для вирішення незбалансованих задач багатокласової класифікації

The object of the study is the classification process of imbalanced multi-class data in air quality analysis based on the air pollution standard index (ISPU), which involves numerical environmental features and categorical output classes. This study addresses the problem of imbalanced multi-class classification in air quality data based on the air pollution standard index (ISPU), where conventional classification techniques tend to be biased toward majority classes and fail to accurately identify minority classes. To overcome this limitation, a cost-sensitive learning strategy combined with adaptive error weighting, grid search with k-fold cross-validation, and principal component analysis (PCA) is applied. The dataset consists of 1,147 samples with an imbalanced distribution across three classes. The results demonstrate that the proposed method achieves an accuracy and F1-score of 98.55% and an area under the ROC curve (AUC) of 0.999, while significantly improving minority class sensitivity. This performance is explained by the cost-sensitive method that increases the penalty for minority class errors and by PCA, which enhances feature representation and learning stability. Compared to existing methods, the proposed method provides a more balanced and robust classification performance without modifying the original data distribution. This method can be effectively applied to ISPU-based air quality classification results and other imbalanced multi-class classification problems, although it requires careful parameter optimization for different data characteristics.