Mislabeling (label noise) - an often overlooked issue in chemometrics

In supervised learning, errors can arise in label values. This fact is well known in machine learning (ML), where it has been studied since the late 1960s, but is often ignored in chemometrics. The main reason for mislabeling is human errors when supervision is carried out by a person. It is also inevitable in cases where a continuous process is divided into stages for their subsequent supervised discrimination. Another reason is the imperfection of an instrumental method, which poorly reflects the differences between classes.

This paper pursues two goals. The first one is to introduce the concept of mislabeling in chemometrics. The second goal is to share our experience in solving mislabeling problems that arise in real-world examples. In our opinion, the most appropriate approach is the integration of the domain knowledge with the outlier detection methods, which should be based on probabilistic (soft) classifiers. This approach is used systematically and is illustrated by three cases that employ both well-established and new methods.

Mislabeling is a real problem that can occur across a variety of applications, reaching up to 20%, and in some areas, such as medicine, this rate can be even higher. There is no doubt that mislabeling deserves investigation and correction. In mission-critical fields, particularly in medicine, the cost of an error caused by mislabeled data can be significant, measured not only in financial losses but also in human lives.