Title: Robust kernel PCA by weighting observations
Authors: Lauri Heinonen - University of Turku (Finland) [presenting]
Joni Virta - University of Turku (Finland)
Abstract: A robust version of kernel principal component analysis is presented, which tolerates outliers. Observations are weighted by iteratively calculating the mean and covariance matrix in feature space, calculating a weight using a function of the Mahalanobis distance and using that in calculating the next mean and covariance. The method is closely connected to classical $k$-step $M$-estimates. All calculations are done in the feature space with the kernel matrix. The convergence of the weights is discussed. The results are illustrated with examples and compared to other relevant methods.