基于鉴别的分析类不平衡学习.docx

AbstractFeature extraction is an important research topic in the field of pattern recognition. Theclass-specific idea tends to recast a traditional multi-class feature extraction and recognition taskinto several binary class problems. In this way, the size of specific class is much smaller than all theother classes thus the problem of class imbalanace occurs. Traditional Fisher Discrminant Analysis(FDA) aims to maximize the between class scatter and minimize the within class scatter, which isdeveloped based on the assumption that samples from two classes are subjected to Gaussiandistributions. We propose three approaches to solve the problem of class imbalance.We first propose an approach called class-balanced discrimination (CBD). For a specific class,we select a reduced counterpart class whose data are nearest to the data of specific class, and furtherdivide them into smaller subsets, each of which has the same size as the specific class, to achievebalance. Then, each subset is combined with the minority class, and linear discriminant analysis(LDA) is performed on them to extract discriminative vectors. To further remove redundantinformation, we impose orthogonal constraint on the extracted discriminant vectors amongcorrelated classes which is called orthogonal balanced discrimination (OCBD). In order to get a sortof discriminant features uncorrelated with each other, we also propose a method called uncorrelatedbalanced discrimination (UCBD).We then propose a new class-balanced discrimination based on active learning (ALCBD). Wedesign two strategies to choose most distinctive subset from neighbor sample set which consider thetotal scatter and discriminability, respectively. The new balanced subset is combined with theminority class, and LDA is performed on them to extract discriminative vector.At last, we propose a kernel class-balanced discrimination (KCBD). To solve theclass-imbalance issue in nonlinear subspace, we map the data from input subspace to a highdime