Identifying Mislabeled Training Data

摘要

This paper presents a new approach to identifying and eliminating mislabeledtraining instances for supervised learning. The goal of this approach is toimprove classification accuracies produced by learning algorithms by improvingthe quality of the training data. Our approach uses a set of learningalgorithms to create classifiers that serve as noise filters for the trainingdata. We evaluate single algorithm, majority vote and consensus filters on fivedatasets that are prone to labeling errors. Our experiments illustrate thatfiltering significantly improves classification accuracy for noise levels up to30 percent. An analytical and empirical evaluation of the precision of ourapproach shows that consensus filters are conservative at throwing away gooddata at the expense of retaining bad data and that majority filters are betterat detecting bad data at the expense of throwing away good data. This suggeststhat for situations in which there is a paucity of data, consensus filters arepreferable, whereas majority vote filters are preferable for situations with anabundance of data.