Research on Classification Label Denoising Algorithm Based on Granular Ball

This paper presents a granular ball denoising method (GBD) which can effectively improve the accuracy and robustness of classification algorithm. GBD method first uses the self-adaptive hypersphere to cover the data space, then eliminates the data not contained in the hypersphere and the noise data in the sphere, and finally uses only the hypersphere data for training, so as to reduce the sample data and improve the data quality. Experiments show that using the data obtained by this method for training can greatly improve the accuracy of the classification algorithm. In addition, because the hypersphere has good adaptive ability, GBD also has excellent robustness. The experiment shows that GBD sampling data training can still get good results after adding noise to the data set. Therefore, GBD is an efficient and robust denoising method.