Use of Instance Typicality for Efficient Detection of Outliers with Neural Network Classifiers

Abstract

Detection of outliers is one of the data pre-processing tasks. In all the applications, outliers need to be detected to enhance the accuracy of the classifiers. Several different techniques, such as statistical, distance-based and deviation-based outlier detection exist to detect outliers. Many of these techniques use filter method. A wrapper method using the concept of instance typicality may also be used to detect outliers. This paper deals with a new wrapper method that builds an initial model using neural networks and treats values at the output of neurons in the output layer as the typicality scores. Instances with lowest output values are treated as potential outliers. In addition, the method is also useful to build compact and accurate classifiers by selecting a few most typical instances resulting in significant reduction in storage space. The method is generic and thus can also be used for instance selection with any kind of classifiers. Resultant compact models are useful for imputation of missing values.