Feature Selection For Cost Sensitive Via Minimizing Localized Generalization Error Model

The objective of the classical classifier is minimizing the classification error. The cost ofthe misclassification for all samples is the same. In cost sensitive learning, misclassificationof samples in different classes will lead to different costs. This situation occurs in many realapplications, for examples, medical detection, face recognition, credit detection, etc. Costsensitive learning has been studied for many years but feature selection for cost sensitivelearning is seldom discussed.Feature selection is widely used in pattern recognition and machine learning tasks. Theclassification problem can be described more precisely by selecting relevant features. TheElimination of redundant features enhances the efficiency of classifier learning. Featureselection for cost sensitive learning not only reduces the cost of data collection, but alsoimproves the performance of the classifier and minimizes the misclassification cost.The Localized Generalization Error Model (L-GEM) evaluates the generalizationcapability of a classifier based on its training error and the stochastic sensitivity, which hasbeen used in the architecture selection of classifiers. In this work, a cost sensitive localizedgeneralization error model (CS-LGEM) based feature selection method is proposed. Twotypes of feature elimination algorithms are presented and the CS-LGEM model is utilized toevaluation the fitness of the set of features. The training phase of the RBFNN is improvedwith cost sensitive factors which minimizes the misclassification cost.As a case study, the cost sensitive steganalysis of JPEG images is discussed in this work.Steganalysis detects whether hidden message is encrypted in an image file, which is a hottopic in network security researches. The cost of misclassifying an image with malicioushidden message as a normal image could cost a lot of damage. So, the cost sensitive problemis important to steganalysis researches. Experimental results show that the proposed methodoutperforms other algorithms.