Supervised Hyperparameter Estimation for Anomaly Detection

Abstract

The detection of anomalies, i.e. of those points found in a dataset but which do not seem to be generated by the underlying distribution, is crucial in machine learning. Their presence is likely to make model predictions not as accurate as we would like; thus, they should be identified before any model is built which, in turn, may require the optimal selection of the detector hyperparameters. However, the unsupervised nature of this problem makes that task not easy. In this work, we propose a new estimator composed by an anomaly detector followed by a supervised model; we can take then advantage of this second model to transform model estimation into a supervised problem and, as a consequence, the estimation of the detector hyperparameters can be done in a supervised setting. We shall apply these ideas to optimally hyperparametrize four different anomaly detectors, namely, Robust Covariance, Local Outlier Factor, Isolation Forests and One-class Support Vector Machines, over different classification and regression problems. We will also experimentally show the usefulness of our proposal to estimate in an objective and automatic way the best detector hyperparameters.

Comparative Analysis of Outlier Detection Approaches in MetOcean Domain: Extreme Value Analysis, Functional Data Analysis, and Unsupervised Learning

“…In our study, we found that to build the UL models for outlier detection require values for the contamination parameter to effectively detect outliers within the dataset. This parameter denotes the proportion of outliers present in the dataset and is relevant for both supervised (Bella, Fernández, & Dorronsoro, 2020) and unsupervised (Z. Xu, Kakde, & Chaudhuri, 2019) models.…”