B0710
Title: Trimming outliers in matrix-variate normal mixtures using the OCLUST algorithm
Authors: Katharine Clark - McMaster University (Canada) [presenting]
Paul McNicholas - McMaster University (Canada)
Abstract: The original version of the OCLUST algorithm trims outliers iteratively in multivariate normal mixtures. Leveraging that Mahalanobis squared distances are chi-squared distributed (or scaled beta-distributed when using sample parameter estimates) for multivariate normal data, suspected outliers are removed one by one until the subset log-likelihoods conform to the specified distribution. The OCLUST algorithm is extended to matrix-variate normal mixtures. Using the matrix-variate normal analogue of Mahalanobis squared distance, it is shown that the log-likelihoods approximate a shifted chi-squared mixture distribution. This distribution is simultaneously employed to detect likely outliers in matrix-variate normal mixtures as well as to predict the proportion of outlying points.
