Angle-Based Dictionary Learning for Outlier Detection

We propose an extension of the Angle-Based Outlier Detection (ABOD) technique by combining it with a Dictionary Learning (DL) problem. The ABOD method benefits from this change by constructing an inlier base with the vector atoms obtained from the DL problem. Our method computes the dictionary
$\boldsymbol{D}$
and calculates the angles between the data points in the feature space and the obtained atoms. We show that using a dictionary in the ABOD method can improve the results in anomaly detection tasks. Generally, the atoms can capture the direction of the inliers, thus better isolating the outliers.