Construction of hierarchical classification model for product management: Penalized Information Gain considering dynamic weight coefficients

The object of this study is the process of constructing hierarchical classifiers for textual data within a defined taxonomy. The task addressed focuses on minimizing cascading errors and enhancing classification consistency across all hierarchy levels, a critical challenge for deep and imbalanced hierarchical structures. The proposed model leverages the Penalized Information Gain (PIG) criterion with dynamically adjusted weight coefficients. A model for hierarchical text classification has been built. It aims to improve classification accuracy and preserve the structural logic of data within multi-level taxonomies. Data featuring a multi-level taxonomy that meets classification requirements have been synthesized and are employed for training and testing classifiers. The performance of local and global hierarchical classification models was compared against conventional classifiers that do not account for taxonomic relationships between classes. The results demonstrate that using weight coefficients based on hierarchical levels enables an adequate representation of taxonomic dependences, which is crucial for maintaining data integrity and improving categorization quality at various levels. Experimental findings show an 8 % increase in the F1 score at the class and subclass levels. A distinctive feature of the model built is the integration of dynamic weights into the splitting criterion, which allows hierarchical dependences between classes to be effectively addressed and cascading errors, typical of conventional approaches, to be minimized. The model’s practical application spans product management systems in e-commerce, text analytics in the restaurant business, and automated categorization systems for multi-level structures