MeaTS: An End-to-End Meta-Enhanced Attention for Time-Series Forecasting With Missing Data

Time-series forecasting with missing values remains a critical challenge across diverse domains. While Transformer-based models excel with complete data, they suffer from “attention sink” where attention mechanisms disproportionately focus on missing data points, creating a self-reinforcing feedback loop that degrades performance. To address this, we propose meta-enhanced attention (MeaTS), a novel end-to-end Transformer-based framework specifically designed for robust multivariate time-series forecasting with missing values. Our approach introduces two key innovations, first, an MeaTS mechanism that dynamically adjusts attention weights by suppressing focus on missing values while amplifying attention to observed data, effectively mitigating the attention sink problem; second, an extracting latent value module that transforms data with missing values into informative features through frequency-domain representations, enhancing data representation without explicit imputation. Extensive experiments demonstrate that MeaTS achieves a 19.86% improvement in mean absolute error compared to state-of-the-art models, such as SDformer, GinAR+, and BiTGraph.