Text Anomaly Detection with ARAE-AnoGAN
Submission Type
Event
Faculty Advisor
Mark Liffiton
Expected Graduation Date
2020
Location
Room E102, Center for Natural Sciences, Illinois Wesleyan University
Start Date
4-4-2020 11:30 AM
End Date
4-4-2020 11:45 AM
Disciplines
Computer Sciences | Education
Abstract
Anomaly detection is the identification of events or observations that deviate from the expected behavior. In recent years, there has been extensive research in using deep learning methods to detect anomalies in images, but few have been applied to text data. Deep learning is a technique involving multiple layers of artificial neural networks for computers to discover patterns in data on their own through learning from examples. Successful applications of deep learning include image recognition, recommendation systems, and self-driving cars. In this work, to test the applicability of deep learning to text anomaly detection, we present ARAE-AnoGAN, a semi-supervised learning method that uses an adversarially regularized autoencoder (ARAE) to model discrete tokens in sentences of normal training data. Anomalies are then detected via a combined anomaly score based on the building blocks of the trained model - consisting of an autoencoder reconstruction error and a discriminator feature residual error. Finally, we present experimental results demonstrating the effectiveness of deep learning methods in text anomaly detection.
Text Anomaly Detection with ARAE-AnoGAN
Room E102, Center for Natural Sciences, Illinois Wesleyan University
Anomaly detection is the identification of events or observations that deviate from the expected behavior. In recent years, there has been extensive research in using deep learning methods to detect anomalies in images, but few have been applied to text data. Deep learning is a technique involving multiple layers of artificial neural networks for computers to discover patterns in data on their own through learning from examples. Successful applications of deep learning include image recognition, recommendation systems, and self-driving cars. In this work, to test the applicability of deep learning to text anomaly detection, we present ARAE-AnoGAN, a semi-supervised learning method that uses an adversarially regularized autoencoder (ARAE) to model discrete tokens in sentences of normal training data. Anomalies are then detected via a combined anomaly score based on the building blocks of the trained model - consisting of an autoencoder reconstruction error and a discriminator feature residual error. Finally, we present experimental results demonstrating the effectiveness of deep learning methods in text anomaly detection.