Anomalies in multidimensional contexts

Abstract

This paper investigates the problem of presenting anomalies in a multidimensional data set. In such a data set, some dimensions may be merely descriptive, while others represent measures and attribute values used to determine whether the data is anomalous. A data cube of the descriptive dimensions is used as a data structure to partition the data set into sub-groups at each note, or context. It is shown that it is possible for a datum to be anomalous in more than one context. Previous work has dealt with this problem by embedding exception indicators in the data cube. Since the data cube is potentially large and anomalies are rare, searching for anomalies is inconvenient. Instead, it is proposed to construct a report for each anomaly that shows its status in each possible context. This results in a direct presentation of anomalous data.