AUTOMATIC DETECTION OF OUTLIERS IN MULTIVARIATE DATA

摘要

Automatic detection of outliers in multivariate data is disclosed. One example is a system including a distance generator to determine Mahalanobis distances for data elements of multivariate input data, the Mahalanobis distance of a given data element indicative of a measure of distance of the given data element from a probability distribution associated with the input data. A filter module selects a portion of the input data, the cohesive portion indicative of cohesive data elements, and the selection based on minimizing the Mahalanobis distances, and identifies candidate outliers to be data elements not in the cohesive portion. An outlier detector automatically detects outliers of the candidate outliers, the detection based on median absolute deviations of the Mahalanobis distances of the input data, and a modified z-score. A display module generates a visual representation of the detected outliers.