A reconstruction error-based framework for label noise detection

Abstract Label noise is an important data quality issue that negatively impacts machine learning algorithms. For example, label has been shown to increase the number of instances required train effective predictive models. It also model complexity and decrease interpretability. In addition, can cause classification results a learner be poor. this paper, we detect with three unsupervised learners, namely $$\textit{principal component analysis} \hbox { (PCA)}$$ principal componentanalysis(PCA) , $$\textit{independent (ICA)}$$ xmlns:mml="http://www.w3.org/1998/Math/MathML">independent />(ICA) autoencoders. We evaluate these learners on credit card fraud dataset using multiple levels, then compare traditional Tomek links filter. Our binary approach, which considers as anomalies, uniquely uses reconstruction errors for noisy in order identify filter noise. detecting instances, discovered autoencoder algorithm was top performer (highest recall score 0.90), while performed worst 0.62).