Detection and correction of mislabeled samples based on graph structure

Machine learning trains and obtains learning models based on a large amount of training samples. Mislabeled training samples will affect the generalization/performance of the final predictive model. Some methods of detecting/correcting mislabeled samples such as graphbased methods, are proposed and used in machine learning to improve predictive models' generalization. However, these methods do not perform well for high-dimensional samples. In this paper, we present three algorithms for detecting/correcting mislabeled samples in highdimensional feature space. First, we propose an improved high-dimensional detection algorithm: PCA--RNG. Next, we introduce a notion of -scalar relative neighbourhood graph ( -SRNG) and study its relationship with relative neighbourhood graph (RNG) andrelative neighbourhood graph ( -RNG). Then, we give an alternative high-dimensional detection algorithm: PCA--SRNG. After detecting mislabeled training samples, it is necessary to correct these mislabeled samples. Then we further propose a scalar-adapted correction algorithm: Fat location correction/deletion. Finally, we explore and validate our algorithms based on real datasets with high-dimensional features.