C-GATS: Conditional Generation of Anomalous Time Series

Vikramank Singh, Abishek Sankararaman, Murali Balakrishnan, Zhao Song

Keywords: Time Series, Anomaly Detection, Generative Models, Data Augmentation, Neural Networks

TL;DR: A method to generate synthetic anomalous time series samples with labels to improve performance of deep time series anomaly detection models

Abstract: Sparsity of the data needed to learn about the anomalies is often a key challenge that is faced when it comes to training deep supervised models for the task of Anomaly Detection (AD). Generating synthetic data by applying pre-determined transformations that conform to a set of known invariances has shown to improve performance of such deep models. In this work we present C-GATS to show that one can learn a much larger invariance space using the available sparse data by training a conditional generative model to do Data Augmentation (DA) for anomalous Time Series (TS) in a model-agnostic way. Particularly, we factorize an anomalous TS sequence into 3 attributes— normal sub-sequence, anomalous sub-sequence, and position of the anomaly and model each of them separately. This factorization helps exploit samples from the dominant class i.e normal TS to train a generative model for the sparse class i.e anomalous TS. We provide an exhaustive study to showcase that C-GATS not only learns to generate different types of anomalies (eg: point anomalies and level-shift) but those generated anomalies improve performance of multiple SOTA TS AD models on a set of popular public TS AD benchmark datasets.