Abordagem para redução da hierarquia de classes em problemas de classificação hierárquica

Abstract

This work presents an approach to simplify the complexity in hierarchical multi-label classification problems, where an instance can be categorized into multiple classes organized in a Directed Acyclic Graph (DAG) structure. These problems are commonly encountered in domains such as bioinformatics, image recognition, and text classification, where class granularity and imbalance can lead to computational and prediction accuracy challenges. The project developed a preprocessing method to reduce class hierarchy, aiming to simplify redundancies and complexity. The proposed approach was evaluated through tests on ten hierarchical multi-label classification datasets. After reducing the classes, the datasets were tested using the Clus-HMC hierarchical classifier and compared with results from the literature. Using the Wilcoxon statistical test, it was observed that the reduction in class hierarchy resulted in statistically significant improvements in AUPRC (Area Under the Precision-Recall Curve) values. This indicates that the proposed preprocessing method effectively simplified hierarchical structures, consequently increasing the accuracy and efficiency of hierarchical multi-label classification models.