SAITS: Self-attention-based imputation for time series

Missing data in time series is a pervasive problem that puts obstacles the way of advanced analysis. A popular solution imputation, where fundamental challenge to determine what values should be filled in. This paper proposes SAITS, novel method based on self-attention mechanism for missing value imputation multivariate series. Trained by joint-optimization approach, SAITS learns from weighted combination two diagonally-masked (DMSA) blocks. DMSA explicitly captures both temporal dependencies and feature correlations between steps, which improves accuracy training speed. Meanwhile, weighted-combination design enables dynamically assign weights learned representations blocks according attention map missingness information. Extensive experiments quantitatively qualitatively demonstrate outperforms state-of-the-art methods time-series task efficiently reveal SAITS’ potential improve learning performance pattern recognition models incomplete real world.