engleski

Outlier detection in experimental data using a modified expectation maximization algorithm

The paper studies the problem of clustering data sets with the E-M (Expectation Maximization) algorithm. Within the E-M algorithm is implemented procedure omitting data that have a low probability of belonging to a Gaussian mixture model components. For this purpose, the threshold is determined by the rejection of data, which are considered as outliers. For this procedure is used Mahalanobis distance of the observed data to the expectations component model that describes a particular cluster. Mahalanobis distance in this situation proved to be a good choice for Gaussian mixture models that describe clusters.

E-M algorithm ; Mahalanobis distance ; data clustering ; outliers

nije evidentirano