NLP Based Anomaly Detection for Categorical Time Series

TL;DR

This paper introduces a novel NLP-inspired approach for anomaly detection in multi-dimensional categorical time series, leveraging language modeling techniques to improve detection and root cause analysis.

Contribution

It formalizes an analogy between categorical time series and NLP, developing and testing three machine learning models for anomaly detection and root cause analysis.

Findings

Effective anomaly detection in categorical time series

Improved root cause investigation capabilities

Demonstrated the strength of NLP analogy in this context

Abstract

Identifying anomalies in large multi-dimensional time series is a crucial and difficult task across multiple domains. Few methods exist in the literature that address this task when some of the variables are categorical in nature. We formalize an analogy between categorical time series and classical Natural Language Processing and demonstrate the strength of this analogy for anomaly detection and root cause investigation by implementing and testing three different machine learning anomaly detection and root cause investigation models based upon it.