Canonical Correlation Analysis Method Based On Spectral Regularization Technique Research

CCA is a multivariate statistical analysis method by which the relationship between two sets of variables is studied, and it is widely used in many fields. CCA only focus on the correlation between the paired samples, however, the difference between the classes isn’t fully taken into account, and the difference between the classes is usually the focus of recognition methods. In DCCA the class information is introduced. The extracted feature by using DCCA can make the correlation between samples from same class maximum and the correlation between samples from different classes minimum at the same time. The recognition performance of DCCA is significantly higher than the classic CCA.In practical applications, noise is inevitably in the sampling process. Noise has a great impact on small eigenvalues in the characteristic spectrum space, and can make the square root of the small eigenvalues quite unstable. To make solving becomes stable, regularization technique is introduced into the DCCA method and then ER-DCCA algorithm is proposed. The regularization technique makes the solving with small eigenvalues stable and the recognition performance is improved. Experiments are carried out on a variety of data sets and the effectiveness of ER-DCCA has been verified.On the premise of canonical correlation analysis based on class label, Y-CCA is proposed by introducing class information. In Y-CCA algorithm, the class label of the test sample set is predicted by known information. The algorithm uses the predicted class label to identify. Experiments are carried out on a part of data set from UCI library and the results of these experiments illustrate the feasibility of the Y-CCA algorithm.