Kernel outlier detection

Abstract

new anomaly detection method called kernel outlier detection (KOD) is proposed.
It is designed to address challenges of outlier detection in high-dimensional
settings. The aim is to overcome limitations of existing methods, such as dependence
on distributional assumptions or on hyperparameters that are hard to tune.
KOD starts with a kernel transformation, followed by a projection pursuit approach.
Its novelties include a new ensemble of directions to search over, and a
new way to combine results of different direction types. This provides a flexible
and lightweight approach for outlier detection. Our empirical evaluations illustrate
the effectiveness of KOD on three small datasets with challenging structures,
and on four large benchmark datasets.