Data Editing for Neuro-Fuzzy Classifiers
Bogdan Gabryś · 2001
In this paper we investigate the potential benefits and limitations of various data editing procedures when constructing neuro-fuzzy classifiers based on hyperbox fuzzy sets. There are two major aspects of data editing which we are attempting to exploit: a) removal of outliers and noisy data; and b) reduction of training data size. We show that successful training data editing can result in constructing simpler classifiers (i.e. a classifier with a smaller number and larger hyperboxes) with better generalisation performance. However we also indicate the potential dangers of overediting which can lead to dropping the whole regions of a class and constructing too simple classifiers not able to capture the class boundaries with high enough accuracy. A more flexible approach than the existing data editing techniques based on estimating probabilities used to decide whether a point should be removed from the training set has been proposed. An analysis and graphical interpretations are given for the synthetic, non-trivial, 2-dimensional classification problems.