An Outlier Detection Approach Based on WGAN-Empowered Deep Autoencoder

Modelling normal data is one of the major challenges in outlier detection. Deep learning has been proven to be effective in modelling underlying distributions of input training data. However, the existing deep learning-based methods normally focus on how to alleviate the negative impact from the presence of outliers when they model the full dataset. Besides, insufficient size of training data also leads to unsatisfactory training of deep networks. This paper proposes a WGAN-empowered deep Autoencoder-based Outlier detection approach (GAEO for short) which presents a novel way to directly and effectively model reliable normal data. GAEO first obtains initial normal data by a proposed initial distance-based outlier scoring function, then constructs a WGAN network to perform normal data augmentation to obtain adequate training samples for subsequent normal data modelling. The deep autoencoder is then trained to discover and yield distributions and patterns of normal data, and the reconstruction error is defined as outlierness of each data object. In our experiments, we investigate the effectiveness of our method GAEO compared with three state-of-the-art outlier detectors on ten real-world datasets, and discuss the impact of the parameters in GAEO. We show that our method GAEO significantly outperforms its contenders by 4% to 20% AUC improvement.

I. Introduction

An outlier is a data object which deviates so much from other observations. Outlier detection refers to the problem of finding instances that do not conform to the normal and expected behaviours. The investigation of outlier detection finds wide applications in numerous fields, e.g., network intrusion detection, financial fraud detection, and medical diagnosis. The success of deep neural networks provides broad opportunities for developing new outlier detection techniques to address many practical problems.