CSDI: Conditional Score-based Diffusion Models for Probabilistic Time Series Imputation

The imputation of missing values in time series has many applications in and finance. While autoregressive models are natural candidates for series imputation, score-based diffusion models have recently outperformed counterparts including autoregressive models in many tasks such as generation and audio synthesis, and would be promising for time series. In this paper, we propose Conditional Score-based Diffusion models Imputation (CSDI), a novel time series imputation method that utilizes-based diffusion models conditioned on observed data. Unlike existing-based approaches, the conditional diffusion model is explicitly trained imputation and can exploit correlations between observed values. On and environmental data, CSDI improves by 40-65% over existing imputation methods on popular performance metrics. In addition, imputation by CSDI reduces the error by 5-20% compared to the-of-the-art deterministic imputation methods. Furthermore, CSDI can also applied to time series interpolation and probabilistic forecasting, and is with existing baselines. The code is available at: //github. com/ermongroup/CSDI.