DiffMissing: Denoising Diffusion Model for Multivariate Time Series Forecasting with Variable Missing

Missing values are prevalent in multivariate time series forecasting, especially when certain variables are entirely missing, posing a significant challenge to traditional methods. Two-stage models that combine imputation and forecasting methods are prone to introduce bias and lead to error accumulation. In contrast, existing end-to-end forecasting models with missing data recovery components fail to ensure data consistency when missing variables. Diffusion models, known for their robust generative capabilities, prefer to generate consistent results based on the available observations. Consequently, we propose a novel end-to-end denoising diffusion model named DiffMissing for multivariate time series forecasting with variable missing. DiffMissing employs a contextual conditional encoder to extract local and global contextual information from the observed variables, assisting the denoising network in generating missing variables that align with the true distribution. Meanwhile, the dynamic interaction between observed and missing variables is modeled through a carefully designed adaptive sparse attention, ensuring consistency among variables. Extensive experimental results show that DiffMissing outperforms existing methods in prediction performance on multiple real-world datasets, especially in the high missing rate scenario of 90%, where the MAE is improved by an average of 9.90% over the best baseline.