Deleting or keeping outliers for classifier training?
Antonio J. Tallón‐Ballesteros, José C. Riquelme · 2014
This paper introduces two statistical outlier detection approaches by classes. Experiments on binary and multi-class classification problems reveal that the partial removal of outliers improves significantly one or two performance measures for C4.5 and 1-nearest neighbour classifiers. Also, a taxonomy of problems according to the amount of outliers is proposed.