Discriminant Analysis Based On Under-sampling Technique For Class Imbalance Learning

Biometrics is an authentication technology, which is the first pass of an information securitysystem. Discriminant analysis has been widely applied to many fields, such as image classificationand biometrics. In the practical application, the distribution of the data is usually uneven. The resultwill bias toward the majority class and has a low performance on the minority class when using thetraditional algorithms dealing with class-imbalanced problem. To solve this problem, we propose aseries of feature extraction approaches in the re-balanced sample set based on the idea of “specificclass”.Firstly, we propose a discriminant analysis using under-sampling technique based on fusedsimilarity measurement for class-imbalanced learning(FUDA-CIL). Under-sampling is commonlyused in supervised class-imbalance learning problem. We use the idea of “specific class” and thefused similarity measurement to select the nearest neighboring samples from negative class of theevery sample of the specific class. Then, we extract discriminative vectors from the re-balancedsample set.Secondly, we map the class-imbalanced samples into an implicit higher feature space and thenuse the kernel method to solve the nonlinear distribution problem, proposing a novel algorithmnamed kernel orthogonal discriminant analysis based on under-sampling technique forclass-imbalanced learning (KOUDA-CIL). This method uses the nonlinear space mapping, andimposes orthorgonal constraint on the extracted discriminant vectors among correlated classes toremove redundant information of the vectors.Finally, in order to deal with the diversity distribution problem of input data, we introduce themultiple kernel learning to solve class-imbalanced problem, and propose a multi-kerneldiscriminant analysis based on under-sampling algorithm (MKUDA-CIL). We transfer the inputdata into several different feature spaces, and then map them into different kernel spaces. We obtainthe discriminant vectors in a new combined multi-kernel space. In order to enhance thedisciminability of the projection vectors, we impose uncorrelated constraint to MKUDA, and thenpropose a multi-kernel uncorrelated discriminant analysis based on under-sampling algorithm(MKUUDA-CIL).The experimental results on Coil20, MNIST and YaleB databases demonstrate theeffectiveness of our methods. Compared with some representative methods, our approaches can solve the class imbalance problem more effectively.