Class label versus sample label-based CCA

摘要

When correlating the samples with the corresponding class labels, canonical correlation analysis (CCA) can be used for supervised feature extraction and subsequent classification. Intuitively, different encoding modes for class label can result in different classification performances. However, actually, when the samples in each class share a common class label as in usual cases, a unified formulation of CCA is not only derived naturally, but also more importantly from it, we can get some insight into the shortcoming of the existing feature extraction using CCA for sequent classification: the existing encodings for class label fail to reflect the difference among the samples such as in central region of class and those in mixture overlapping region among classes, consequently resulting in its equivalence to the traditional linear discriminant analysis (LDA) for some commonly-used class-label encodings. To reflect such a difference between the samples, we elaborately design an independent soft label for each sample of each class rather than a common label for all the samples of the same class. A purpose of doing so is to try to promote CCA classification performance. The experiments show that this soft label based CCA is better than or comparable to the original CCA/LDA in terms of the recognition performance.