A study of outliers for robust independent component analysis

The impact of outliers on the signal separation performance of an independent component analysis (ICA) algorithm is an important characteristic in assessing the algorithm´s utility in real-world applications. If an ICA estimator has the property of B-robustness, the influence of an extreme point is bounded, leading to good separation performance in the presence of outliers. In recent work, major ICA estimators, such as FastICA, have been proven not to be B-robust. We seek to enhance the non-B-robust FastICA estimator by the introduction of K-means clustering for outlier mitigation. We compare our algorithm with the B-robust β-divergence algorithm by conducting a simulation to reproduce published results. The paper demonstrates the utility of the K-means clustering algorithm to mitigate a class of outliers such that our ICA separation performance is at least equal to that of published results for the B-robust β-divergence estimator.