Label denoising based on Bayesian aggregation

Abstract

‎Label noise is a common problem that affects supervised learning and can produce misleading results‎. ‎It is shown that only 5% of switched labels lead to a decrease of performances‎. Therefore, the true class of an instance must be distinguished from its observed label‎. ‎In the past decade‎, ‎classification in presence of label noise was the topic of interest‎. ‎Several scholars focused on kNN-based approaches for data cleansing‎. ‎These types of approaches often are susceptible to high label noise rate and when a batch of instances with noisy labels are exist they may deteriorate the results‎. The problem arises since the methods have a local view of instances‎. ‎Another approach is to have a global view of instances‎. In a global view‎, ‎instances with large distance from their respective classes are detected as noisy‎. A potential problem however is the determination of a threshold‎. ‎An inappropriate threshold may lead to detection of a correct instance as noisy instance‎. ‎In this paper a new method for label denoising based on Bayesian aggregation is proposed which solves the problems of kNN-based approaches by aggregating the local and global views of instances‎. ‎The aggregation of local and global information leads to a more robust and accurate detection of instances with noisy labels and estimation of their true labels‎. ‎The experimental results show the capabilities and robustness of the proposed method‎.