A robust novelty detection framework based on ensemble learning

doi:10.1007/s13042-022-01569-9

A robust novelty detection framework based on ensemble learning

2022-05-13

2

PRE

AI

DOI：10.1007/s13042-022-01569-9

Original

Original request for help

Save

Abstract

En 中文

Novelty detection techniques have been used extensively to discover interesting data patterns in practical applications. However, real-world training data are often contaminated by unknown anomalous examples (outliers), leading to deteriorated detectors. In order to alleviate this problem, this paper proposes a robust novelty detection framework based on ensemble learning. In contrast to traditional parallel outlier ensembles based on variance-reduction, both bias and variance are considered in our ensemble framework. Specifically, we aim to reduce the bias induced by unknown outliers with an iterative mechanism. A weighting scheme is used to combine the result of current iteration with the previous iteration. By gradually removing outliers in the training set, performance of the detector can be improved. In addition, base detectors at all iterations will be aggregated by the weighting scheme in order to realize the variance reduction. Moreover, a flexible function that provides reference ground-truth is proposed so that our detection framework can be effective on different types of data sets. We conduct experiments on 15 benchmark data sets to verify the superiority over parallel ensembles and single models. A case study concerning wind tunnel is also carried out on 10 data sets from a real-world wind tunnel system. Experimental results have shown the superiority of our detector over several competitors.

Keywords:

Novelty detection

Robustness

Ensemble learning

One-class classification

A robust novelty detection framework based on ensemble learning

Abstract

Journal

Organization