On the Impact of Noisy Labels on Supervised Classification Models.

The amount of data generated daily grows tremendously in virtually all domains of science and industry, and its efficient storage, processing and analysis pose significant practical challenges nowadays. To automate the process of extracting useful insights from raw data, numerous supervised machine learning algorithms have been researched so far. They benefit from annotated training sets which are fed to the training routine which elaborates a model that is further deployed for a specific task. The process of capturing real-world data may lead to acquring noisy observations, ultimately affecting the models trained from such data. The impact of the label noise is, however, under-researched, and the robustness of classic learners against such noise remains unclear. We tackle this research gap and not only thoroughly investigate the classification capabilities of an array of widely-adopted machine learning models over a variety of contamination scenarios, but also suggest new metrics that could be utilized to quantify such models’ robustness. Our extensive computational experiments shed more light on the impact of training set contamination on the operational behavior of supervised learners.