RDMI: Recursive Training-Based Diffusion Model for Multivariate Time Series Imputation

Abstract

In this paper, we present a novel approach for imputing missing values in multivariate time series using a recursive training-based diffusion model. Our proposed framework incorporates meta-learning, self-conditioning, and recursive training as key components to enhance imputation performance. We evaluate the model on two publicly available real-world datasets and achieve an improvement in RMSE, MAE, CRPS, MAPE, and SMAPE compared to the state-of-the-art model. Additionally, our ablation study confirms that each proposed technique has a meaningful effect on MTS imputation.