Multivariate Time Series Imputation with Signal-Noise Disentangled Graph Propagation

Tianyuan Zhou, Jianwei Wang, Shuo Zhang, Kai Wang, Ying Zhang, Wenjie Zhang

Keywords: Missing Data; Data Imputation; Multivariate Time Series;

Abstract: Missing data are pervasive in real-world multivariate time series, particularly in large-scale, high-frequency systems. Although recent graph-based and transformer-based methods achieve state-of-the-art (SOTA) performance by performing spatial graph propagation or leveraging self-attention mechanisms, they suffer from two key limitations: (1) treating each time series as an indivisible whole, without uncovering its internal temporal dynamics, and (2) relying on linear projections to connect spatial and temporal representations, which insufficiently depicts the complex spatial-temporal interactions. Motivated by the above limitations, we propose GraphTSI, a Graph-based multivariate Time Series Imputation method with signal-noise decomposition, where the signal component captures predictable dynamics and the noise component reflects unpredictable exogenous shocks of time series. To enable robust decomposition, we propose a prediction–subtraction framework where the prediction step progressively estimates predictable signal component, while the subtraction step uses the discrepancy between this estimate and the observed values to extract the exogenous noise component. Furthermore, for effective spatial-temporal interactions, we build an augmented bipartite graph that captures adaptive, non-linear transformation between spatial and temporal dimensions, and propagates signal and noise components through neighboring time series. Extensive experiments across nine datasets from three real-world domains demonstrate the superiority of GraphTSI, with average MAE improvements of 10.273% and 17.580% over graph-based and transformer-based SOTA methods, respectively.

Primary Area: learning on time series and dynamical systems

Submission Number: 8786