Robust deep auto-encoding Gaussian process regression for unsupervised anomaly detection

Unsupervised anomaly detection (AD) is of great importance in both fundamental machine learning researches and industrial applications. Previous approaches have achieved great advance in improving the performance of unsupervised AD model recently. However, there are still some thorny issues unsolved, especially the problem of efficiency degradation when dealing with high-dimensional data and the inability to maintain robustness when dealing with contaminated data, which have not been addressed simultaneously in the existing models. In our work, we propose a novel hybrid unsupervised AD method, which first integrates convolutional auto-encoder and Gaussian process regression to extract features and to remove anomalies from noisy data as well. Our model behaves more effectively at modeling high-dimension data and more robust to variation of the anomaly rate in dataset. We evaluate its performance on four publicly benchmark datasets and show the state-of-the-art performance against competitive methods. (C) 2019 Published by Elsevier B.V.