Spherical projection for improving the accuracy and robustness of ICA or BSS algorithms against outliers
首发时间:2008-06-16
Abstract:Outliers are a few erroneous points or spiky noise in observed data that usually affect the separation accuracy of ICA or BSS algorithms. This paper proposes spherical projection of outlier data points to improve the accuracy and robustness of such algorithms. The idea is built on the geometrical interpretation of the ICA or BSS model where each data point represents a candidate direction for determining an independent component in the separation process. Its weight in a candidate direction depends on its magnitude projected in that direction. As a result projecting the outliers onto a spherical surface of smaller radius will mitigate the outlier’s weight for representing that direction so as to improve accuracy in finding the correct direction. Separation experiments show that this method can significantly improve the separation result for both sub- and supergaussian sources. As projecting only a few outliers costs only a tiny amount of calculation, it is beneficial to incorporate it into any existing algorithms that are prone to outliers.
keywords: Blind source separation, independent component analysis, outlier data, spherical projection
点击查看论文中文信息
Spherical projection for improving the accuracy and robustness of ICA or BSS algorithms against outliers
摘要:Outliers are a few erroneous points or spiky noise in observed data that usually affect the separation accuracy of ICA or BSS algorithms. This paper proposes spherical projection of outlier data points to improve the accuracy and robustness of such algorithms. The idea is built on the geometrical interpretation of the ICA or BSS model where each data point represents a candidate direction for determining an independent component in the separation process. Its weight in a candidate direction depends on its magnitude projected in that direction. As a result projecting the outliers onto a spherical surface of smaller radius will mitigate the outlier’s weight for representing that direction so as to improve accuracy in finding the correct direction. Separation experiments show that this method can significantly improve the separation result for both sub- and supergaussian sources. As projecting only a few outliers costs only a tiny amount of calculation, it is beneficial to incorporate it into any existing algorithms that are prone to outliers.
关键词: Blind source separation, independent component analysis, outlier data, spherical projection
论文图表:
引用
No.2222317725512135****
同行评议
共计0人参与
勘误表
Spherical projection for improving the accuracy and robustness of ICA or BSS algorithms against outliers
评论
全部评论0/1000