Anomaly detection in dynamic networks

TL;DR

This paper introduces oddnet, a novel feature-based statistical method for detecting anomalies in dynamic networks by modeling temporal dependencies with time series techniques, demonstrated on synthetic and real data.

Contribution

The paper presents oddnet, a new statistical approach for network anomaly detection that leverages time series modeling, filling a gap in existing methods.

Findings

Effective detection on synthetic datasets

Successful application to real-world networks

Implemented as an R package

Abstract

Detecting anomalies from a series of temporal networks has many applications, including road accidents in transport networks and suspicious events in social networks. While there are many methods for network anomaly detection, statistical methods are under utilised in this space even though they have a long history and proven capability in handling temporal dependencies. In this paper, we introduce \textit{oddnet}, a feature-based network anomaly detection method that uses time series methods to model temporal dependencies. We demonstrate the effectiveness of oddnet on synthetic and real-world datasets. The R package oddnet implements this algorithm.