A novel covering rough set model based on granular-ball computing for data with label noise

Abstract

As a novel granular computing model, granular-ball computing (GBC) has a notable advantage of robustness. Inspired by GBC, a granular-ball covering rough set (GBCRS) model whose covering is made up of granular-balls (GBs) is proposed. GBCRS is the first covering rough set that fits the data distribution well. Inheriting the robustness of GBC, GBCRS can work in label noise environments. First, the optimization objective function of GBs in GBCRS is given. In order to ensure the quality of generated GBs, this function is subject to three constraints. Second, the GBCRS model is proposed. The purity threshold is used to relax the related notions so that GBCRS can be used in label noise environments. Subsequently, GBCRS is applied to the covering granular reduction and attribute reduction in label noise environments. In covering granular reduction, we propose an intuitive, understandable and anti-noise GBCRS-based granular reduction (GBCRS-GR) algorithm, which also solves the optimization objective function of GBs. Based on GBCRS-GR, a GBCRS-based attribute reduction (GBCRS-AR) algorithm is proposed with the classification ability of the attribute subset as the evaluation. The experiments on UCI datasets illustrate that proposed algorithm is more robust against label noise than the comparison ones.