A Novel Correntropy Analysis Method with Application to Multi-view Feature Representation

In this paper, a novel correntropy analysis (CORA) method is proposed for multi-view feature representation. By joint utilization the correntropy and nonlinear kernel transformation tools, the presented CORA method is able to measure the localized similarity between two random variables and further reveal the intrinsic relation between them effectively, leading to a high quality feature representation. Unlike many existing techniques for feature representation such as canonical correlation analysis (CCA) and kernel CCA (KCCA), CORA indicates and explores the mutual relation of two random variables according to the probability density. In addition, different from the kernel entropy component analysis (KECA) method revealing the structural information only from a single data space, CORA is able to explore the mutual structural information between two data spaces jointly instead. The effectiveness of the proposed method is evaluated through experiments on audio emotion recognition and face recognition examples. Comparisons are conducted on the statistics machine learning (SML) and deep neural network (DNN) based algorithms. The results show that the proposed CORA method outperforms other methods.