The processing for label noise based on attribute reduction and two-step method

Classification is a mainstream task in machine learning. To achieve good classification results, there are many aspects to consider. Among them, label noise is the most direct and fundamental problem. Nowadays research targets the processing of label noise in numerous aspects, including correction, filtering and enhanced robustness methods. All these methods have improved the classification accuracy to some extent. However, the above studies consider only one approach to label noise, such as solely focusing on filtering or exclusively on correction. Label noise is complex and it is singular to consider only one method to deal with it. For example, contaminated data in a certain class and noise belonging to this class, both belong to the label noise problems, but with completely different distributions and treatments. This requires us to discuss the situations separately and to propose different processes. In this paper, we take this into account and propose a noise processing method that combines revision and filtration (RF). The RF method can follow the different distributions of label noise and perform targeted processes, which is more effective and comprehensive. It can maintain the original data distribution and remove noise as much as possible. On the other hand, high-dimensional datasets are encountered when dealing with label noise. The attribute values of the dataset will be abnormal due to the presence of label noise. Therefore, we suggest an attribute reduction method for the case when label noise exists. The advantage is that it not only removes redundant attributes, but also eliminates attributes interfered with by noise, which is suitable for high-dimensional data with label noise. Experiments prove that our proposed RF algorithm is effective among three classifiers with multiple comparison algorithms. Performing attribute reduction also improves classification accuracy significantly.