Outlier Treatment for SLFNs in Classification

Abstract

In past decades, the single-hidden layer feedforward neural networks (SLFNs) have been frequently used to solve the classification problem. It can form decision regions with arbitrary shapes if activation functions of hidden nodes are chosen properly. However, in data collection and analysis there often exist outliers which affect the performance of classification. In order to enhance the classification performance of the SLFNs, it is important to detect and eliminate these outliers. In this paper, we propose an approach for outlier reduction based on distribution of every feature, in which scores are assigned to patterns. Patterns detected as outliers based on these scores will be eliminated from data set. One interesting observation is that, our approach can obtain high accuracy with fast learning speed if the training set exist patterns deviating from mainstream of the remaining of the data set.