ADDRESSING CLASS IMBALANCE IN HIERARCHICAL MULTILABEL CLASSIFICATION WITH NODE-WISE WEIGHTING

Abstract

Class imbalance is a persistent challenge in hierarchical multilabel (HML) classification tasks where labels form parent-child relationships and samples contain more than one label. It hampers learning as the majority classes dominate training while minority classes are neglected, which is amplified due to hierarchical dependencies. We introduce an adapted node-wise weighting method that enforces hierarchical constraints and reimagine HML imbalance as defined by node frequencies in a dataset. We benchmark it on several HML datasets including functional genomics datasets and an oceanographic dataset. We analyze which evaluation metrics provide a comprehensive assessment that is specifically suited for assessing HML tasks. The experimental results show that our node-wise weighting method consistently improves recall for minority classes without sacrificing much precision on majority classes, outperforming prior techniques for handling class imbalance. These findings show the potential of our method to address class imbalance in HML settings applicable to diverse real-world HML tasks.