Evolving Fuzzy Min-max Neural Network for Outlier Detection

Abstract

Outlier detection is a complex task to perform because of the uncertainty involved in it. Fuzzy logic is more suitable for handling uncertainty. Many applications require real time outlier detection. Neural networks are good at real time operation, online adaption and efficient as they are massively parallel in nature. The hybridization of fuzzy and neural computing system is very promising, since they exactly tackle the situation associated with outliers. In this paper, a Fuzzy min-max neural network is used for outlier detection. In testing phase, a method is proposed for outlier detection which is based on majority voting. User has to define a threshold (t) and if the fuzzy membership value of test pattern in a hyper-box is below t then the pattern will be declared as an outlier with respect to the hyper-box. User should also define a parameter p which decides the percentage of hyper-boxes to be considered for voting a test pattern as an outlier. Experimentation is done on synthetic data and a standard database available on UCI Machine Learning Repository [19]. The proposed method has increased the recognition accuracy whereas the drawback is recall time increased as one more level of voting calculation with a serial time complexity of o(k) is added in the testing phase.