A supervised independent component analysis maximizing distances between features of different classes

摘要

Recently, Independent Component Analysis (ICA) has been applied to not only problems of blind signal separation, but also feature extraction of patterns. However, the effectiveness of pattern features extracted by conventional ICA algorithms greatly depends on pattern sets; that is, how patterns are distributed in the feature space. As one of the reasons, we have pointed out that conventional ICA features are obtained by increasing only their independence even if class information is available. In this context, we can expect that more high-performance features can be obtained by introducing class information into conventional ICA algorithms. In this paper, we propose a supervised ICA (SICA) algorithm that maximizes Mahalanobis distances between features of different classes as well as maximize their independence. In the simulation, the performance of the proposed SICA algorithm is evaluated using three data sets of UCI Machine Learning Repository. We demonstrate that the better recognition accuracy for these data sets is obtained using our proposed SICA. Furthermore, we show that pattern features extracted by SICA are better than those extracted by only maximizing the Mahalanobis distances.