Graph-Guided Reconstruction Diffusion for Multivariate Time Series Anomaly Detection

ICLR 2026 Conference Submission19570 Authors

Keywords: time series anomaly detection, diffusion model, multi-resolution

TL;DR: We introduce GGRD, a graph-guided diffusion model that smoothly captures multi-resolution dynamics and reconstructs signals in a single step, delivering fast and robust anomaly detection for complex multivariate time series.

Abstract: Time series anomaly detection often faces challenges such as non-stationarity and trendiness. Recently, unsupervised learning methods combined with generative models have shown promising prospects in this field, especially the application of multi-resolution technology in anomaly detection has achieved certain results. However, existing models usually ignore the correlations among different features in time series data and the rich multi-resolutional knowledge contained in the original data. To solve this problem, this paper proposes a new Model, \textbf{G}raph \textbf{G}uided \textbf{R}econstruction \textbf{D}iffusion Model (GGRD). GGRD is an end-to-end unsupervised anomaly detection model based on reconstruction. It adopts overlapping sliding Windows to sample multi-resolution data and integrates the similarity prior in the data into the \textbf{G}raph-\textbf{G}uided \textbf{A}ttention (GGA) mechanism, thereby effectively dealing with complex characteristics such as non-stationarity and cross-variable correlations of time series. The experimental results show that GGRD significantly outperforms the existing methods on multiple real datasets. Code is available at \url{https://anonymous.4open.science/r/GGRD-806F/}.

Primary Area: learning on time series and dynamical systems

Submission Number: 19570