GANs for Anomaly Detection on Telco multivariate time series

It's time for a new data science meetup! We will have a meetup on anomaly detection using GANs on March 1.

Presenter: Sara Malacarne, Telenor Research

Abstract:
Anomaly detection is the process of identifying interesting events that deviate from the data’s “normal” behaviour and has many important applications to real case scenarios. In the telecommunications domain, efficient and accurate anomaly detection is vital to be able to continuously monitor the network infrastructure's key performance indicators (KPIs) and alert for possible incidents in time.

Network KPIs are in the form of multivariate time series which, for costs reasons, are not labelled. The main challenges for performing anomaly detection on network data are the following: 1) it is an unsupervised learning problem, 2) temporal and feature-wise correlations have to be exploited in order to reduce false positives, 3) anomalies are not necessarily rare events in the data, 4) the data is high-dimensional.

This work is a first attempt to simultaneously address the first three challenges listed above, with the use of a novel Generative Adversarial Network (GAN), called RegGAN. GANs present in the literature -- such as MAD-GAN, BeatGAN, TadGAN -- have serious drawbacks on highly contaminated data, that is, data with frequent abnormal events. Thus, RegGAN was specifically built to overcome this issue, and it has proven to be robust to contamination experiments performed on open benchmark datasets.