Jason Van Hulse,
Taghi M. Khoshgoftaar () and
Amri Napolitano Additional contact information Jason Van Hulse: Department of Computer and Electrical Engineering and Computer Science, Florida Atlantic University, Boca Raton, FL, USA
Taghi M. Khoshgoftaar: Department of Computer and Electrical Engineering and Computer Science, Florida Atlantic University, Boca Raton, FL, USA
Amri Napolitano: Department of Computer and Electrical Engineering and Computer Science, Florida Atlantic University, Boca Raton, FL, USA
Learning from imbalanced training data can be a difficult endeavour, and the task is made even more challenging if the data is of low quality or the size of the training dataset is small. Data sampling is a commonly used method for improving learner performance when data is imbalanced. However, little effort has been put forth to investigate the performance of data sampling techniques when data is both noisy and imbalanced. In this work, we present a comprehensive empirical investigation of the impact of changes in four training dataset characteristics — dataset size, class distribution, noise level and noise distribution — on data sampling techniques. We present the performance of four common data sampling techniques using 11 learning algorithms. The results, which are based on an extensive suite of experiments for which over 15 million models were trained and evaluated, show that: (1) even for relatively clean datasets, class imbalance can still hurt learner performance, (2) data sampling, however, may not improve performance for relatively clean but imbalanced datasets, (3) data sampling can be very effective at dealing with the combined problems of noise and imbalance, (4) both the level and distribution of class noise among the classes are important, as either factor alone does not cause a significant impact, (5) when sampling does improve the learners (i.e. for noisy and imbalanced datasets), RUS and SMOTE are the most effective at improving the AUC, while SMOTE performed well relative to the F-measure, (6) there are significant differences in the empirical results depending on the performance measure used, and hence it is important to consider multiple metrics in this type of analysis, and (7) data sampling rarely hurt the AUC, but only significantly improved performance when data was at least moderately skewed or noisy, while for the F-measure, data sampling often resulted in significantly worse performance when applied to slightly skewed or noisy datasets, but did improve performance when data was either severely noisy or skewed, or contained moderate levels of both noise and imbalance.