Parameter-Free Outlier Scoring Algorithm Using the Acute Angle Order Difference Distance

摘要

An anomaly scoring algorithm assigns the anomalous rating to an instance in a dataset which provides a large value for an outlier. In 2013, one of the parameter-free techniques called the Ordered Difference Distance Outlier Factor algorithm is proposed. It calculates a score using an ordered difference distance among all instances which derives from the distance matrix sorted in each row before computing the difference. The score contribution from other instances must be compared with the global minimum distance to avoid mis-detecting boundaries. However, this degrades the performance of the detection rate. To avoid the use of the global minimum distance term, the new technique is proposed using the ordered difference distance along the appropriate direction based on the acute angle. This technique is called the acute angle ordered difference distance outlier factor (AOF) algorithm. Three collections of ten synthesized datasets are designed to show the performance of AOF. The AOF algorithm reports very high scores for anomalies in synthetic datasets and has better performance than OOF when the anomalies are close together.