Multi-label learning based on neighborhood rough set label-specific features

Abstract

Multi-label learning emerges as a novel paradigm harnessing diverse semantic datasets. Its objective involves eliciting a prognostic framework capable of allocating correlated labels to an unseen instance. Within the multifaceted domain of multi-label learning, the adoption of a label-specific feature methodology is prevalent. This approach entails the induction of a classification model that forecasts the relevance of each class label, utilizing tailored features specific to each label rather than relying on the original features. However, some irrelevant or redundant features will inevitably be generated when constructing features. To address this issue, we extend the current approach and introduce a straightforward yet potent multi-label learning method named NRS-LIFT, i.e., Neighborhood Rough Set Label-specIfic FeaTures. Specifically, a sample selection method is used to reduce the computational complexity, and then a set of tailored features is customized for each label through the neighborhood rough set. Finally, a learning model is induced to predict unseen instances. To fully evaluate the effectiveness of NRS-LIFT, we conduct extensive experiments on 12 multi-label datasets. Compared with mature multi-label learning methods, it is verified that NRS-LIFT has strong performance for multi-label datasets.