Reproduce “Multivariate Time Series Anomaly Detection via Dynamic Graph Forecasting” on OpenResearch

About this paper

Katrina Chen, Mingbin Feng, Tony S. Wirjanto

The University of Waterloo researchers developed DyGraphAD, a framework for multivariate time series anomaly detection that forecasts dynamic graph structures to capture evolving inter-series relationships and combines this with time series forecasting. The system achieved F1-scores up to 96.55% on benchmark datasets, outperforming existing graph-based anomaly detection methods by better identifying both relational and temporal anomalies.

Key insights

• Anomalies in multivariate time series often manifest as abnormal changes in inter-series relationships over time, which can be effectively identified by forecasting the evolution of dynamic graph structures.
• Combining forecasting errors from both dynamic graph prediction and traditional time series prediction provides a more robust anomaly score, as each component detects different facets of anomalies (relational vs. fine-temporal).
• The use of Dynamic Time Warping for constructing dynamic graphs and a Transformer architecture for modeling graph sequences enables the system to capture nuanced, time-shifted inter-series dependencies.
• Integrating a global static graph with dynamic graphs helps balance the capture of long-term stable relationships and short-term evolving patterns.

Summary by alphaXiv.