Fuzzy Analysis and Classification of Mislabeled and Noisy Data

摘要

A classifier learns from a "training" data set so it can later correctly classify a new pattern from the same population as the training set. However, when the examples for a learning algorithm consist of real world data then they are usually tainted with noise, ambiguity, uncertainty, imprecision, vagueness or incompleteness. Noise may be introduced by outliers; they are the result of some bad measurements or pattern mislabeling. Clearly, classification of such noisy data must be highly efficient and accurate. In this paper, we overcome this problem by introducing an efficient tool for feature selection where "bad" (non-discriminating) features are dropped and "good" features are weighted according to how well they separate classes in a data set. Good features are responsible for "electing" a class that the feature vector under test should naturally belong to. Thus, we call our new method "EFCLASS" denoting Election Fuzzy Classification. The proposed method is simple, fast and accurate. Various data sets that are known to be good examples for a classification algorithm are used to test the performance of the proposed method for the fuzzy classifier.