VAE-GAN based zero-shot outlier detection

Document Type

Conference Proceeding

Source of Publication

ACM International Conference Proceeding Series

Publication Date

11-17-2020

Abstract

© 2020 ACM. Outlier detection is one of the main fields in machine learning and it has been growing rapidly due to its wide range of applications. In the last few years, deep learning-based methods have outperformed machine learning and handcrafted outlier detection techniques, and our method is no different. We present a new twist to generative models which leverages variational autoencoders as a source for uniform distributions which can be used to separate the inliers from the outliers. Both the generative and adversarial parts of the model are used to obtain three main losses (Reconstruction loss, KL-divergence, Discriminative loss) which in return are wrapped with a one-class SVM which is used to make the predictions. We evaluated our method against several datasets both for images and tabular data and it has shown great results for the zero-shot outlier detection problem and was able to easily generalize it for supervised outlier detection tasks on which the performance has increased. For comparison, we evaluated our method against several of the common outlier detection techniques such as DBSCAN-based outlier detection, GMM, K-means and one class SVM directly, and we have outperformed all of them on all datasets.

ISBN

9781450388894

Publisher

ACM

Disciplines

Computer Sciences

Keywords

Deep Learning, Generative Models, Machine Learning, Outlier Detection

Scopus ID

85101692127

Indexed in Scopus

yes

Open Access

no

Link to Full Text

Share

COinS