Correntropy Based Feature Extraction Method For Novelty Detection

Feature extraction is an important research issue in pattern recognition and machine learning. The aim of feature extraction is to overcome the "curse of dimensionality" problem encountered by machine learning. So far, feature extraction methods for tackling the two-class and multi-class problems have been completely developed in domestic and foreign. Similar to the two-class and multi-class cases, novelty detection has to face the curse of dimensionality for tackling high-dimensional data sets. Therefore, the effectiveness of feature extraction is regarded as an important role for novelty detection to deal with high-dimensional data sets.In the dissertation, a novel feature extraction method based on regularized correntropy criterion (FEND-RCC) is proposed for novelty detection. FEND tries to find a subspace by maximizing the difference between the sum of squared deviations of the normal data with their mean and the sum of squared deviations of the novel data with the mean of normal data. In FEND-RCC, the presented criterion aims to maximize the deference between the correntropy of the normal data with their mean and the correntropy of the novel data with the mean of the normal data. Moreover, an L2norm based regularization term is introduced into the objective function of the proposed method, which makes the proposed feature extraction method more robust. The optimal projection vectors of FEND-RCC can be iteratively obtained by the half-quadratic optimization technique. Experimental results on two synthetic data sets and thirteen benchmark data sets for novelty detection demonstrate that FEND-RCC is superior to its related approaches.