Multi-Label Streaming Feature Selection via Class-Imbalance Aware Rough Set

Abstract

Multi-label feature selection aims to select discriminative attributes in multi-label scenario, but most of existing multi-label feature selection methods fail to consider streaming features, i.e. features gradually flow one by one, which is more common in real-world applications. In addition, though there are already some representative works on multi-label streaming feature selection, they fail to tackle the class-imbalance problem, which exists widely in multi-label learning. In fact, class-imbalance will lead to the performance degradation of multi-label learning models. Thus considering class-imbalance problem in multi-label scenario is beneficial to multi-label feature selection because more precise feature evaluation is achieved. Motivated by this, we propose a new rough set named as class-imbalance aware rough set model which can fit class-imbalance problem well. To address streaming features, we construct a novel streaming feature selection framework called SFSCI(Streaming Feature Selection via Class-Imbalance aware rough set), which contains online irrelevancy discarding and online redundancy reduction. Finally, an empirical study on a series of benchmark data sets demonstrates that the proposed method is superior to other state-of-the-art multi-label feature selection methods, including several multi-label streaming feature selection methods.