DITS: Double Imputation for Enhanced Irregular Multivariate Time Series Classification

Irregularly sampled multivariate time series classification is challenging and widely used in various domains, including health care, biology, and space science. Imputation methods treat irregularly sampled multivariate time series as having missing values and achieve superior performance. However, most of the existing methods impute time series by the correlation within data streams ignoring the correlation across data streams (both correlations are important information of multivariate time series). Few methods exploit both two correlations in a sequential way which will cause sub-optimal results. To address this problem, we propose a novel imputation-reconstruction-prediction model with double imputation. The double imputation network exploits two kinds of correlations in a currency way and achieves excellent performance without architecture changes. We apply a joint-optimization training approach with three learning tasks to fit our imputation-reconstruction-prediction architecture and the model is end-to-end trainable. Experiments on public real-world datasets demonstrate the effectiveness of our method.