CNC-NOS: Class noise cleaning by ensemble filtering and noise scoring

Abstract

Obtaining data in the real world is subject to imperfections and the appearance of noise is a common consequence of such flaws. In classification, class noise will deteriorate the performance of a classifier, as it may severely mislead the model building. Among the strategies emerged to deal with class noise, the most popular is that of filtering. However, instance filtering can be harmful as it may eliminate more examples than necessary or produce loss of information. An ideal option would be relabeling the noisy instances, avoiding losing data, but instance correcting is harder to achieve and may lead to wrong information being introduced in the dataset. For this reason, we advance a new proposal based on an ensemble of noise filters with the goal not only of accurately filtering the mislabeled instances, but also correcting them when possible. A noise score is also applied to support the filtering and relabeling process. The proposal, named CNC-NOS (Class Noise Cleaner with Noise Scoring), is compared against state-of-the-art noise filters and correctors, showing that it is able to deliver a quality training instance set that overcomes the limitations of such techniques, both in terms of classification accuracy and properly treated instances.