Integrated Autoencoder-Level Set Method Outperforms Autoencoder for Novelty Detection

Abstract

Novelty detection (ND), also known as one-class classification or anomaly/outlier detection, has attracted a lot of research interest across a range of applications, where abnormal data are limited or are extremely rare, e.g., credit card [26], mobile phone fraud detection [1], mobile robotics [2], sensor networks[3], rumor detection [4], [5], video surveillance [6]–[8] and healthcare [9]–[11] areas. ND aim to train the detectors with the target (normal, negative) class and then identify the deviated data as novelties (abnormal, positive) class using the trained detectors.