A Bayesian classification framework with label corrections

Abstract

The use of unlabeled data is very important for regression and classification analysis in many cases. However, the data may have an extra layer of complexity with some wrongly labelled data points. The traditional semisupervised analysis doesn’t have the mechanism to treat unlabeled data and mislabeled data at the same time. Here, we propose a framework with a Bayesian approach to deal with unlabeled and mislabeled data simultaneously with an extra layer of modeling. The same framework not only works on Gaussian mixture models, but it’s also universally applicable on top of any parametric or non-parametric method, such as the kernel method and Dirichlet Process (DP) priors. With a thorough study of the kernel and Dirichlet Process method, we successfully applied our framework onto these non-parametric methods and achieved satisfactory results in simulations. This work shows the power of our Bayesian framework to solve complex uncertainty in the data structure using non-parametric approaches.