MissTime: State-Aware Retrieval for General Time Series Analysis with Missing Variables

Time series analysis is crucial in a wide range of real-world applications. However, multivariate time series often contain missing variables in practice due to sensor or communication failures, disrupting channel-wise correlations and degrading performance across diverse tasks. Existing approaches use imputation models or retrieval of historical instances as additional context to reconstruct the missing variables. However, we identify that they still face two key challenges: (i) reliance on the imputation model and (ii) retrieval ambiguity when incorporating historical references. Motivated by this, we propose MissTime, a direct retrieval framework that identifies similar historical instances and reconstructs the missing variables with the retrieved full observations. MissTime introduces two key components: (i) Full Historical Bank Collection, which curates a historical time series bank with fully observed variables, providing the basis for data reconstruction; and (ii) State-Aware Retrieval, which extracts underlying state representations from both time series with missing variables and fully observed time series. It then aligns samples with similar states through a contrastive disambiguation objective, mitigating retrieval ambiguity caused by partially observed variables. We conduct extensive experiments on time series benchmark datasets demonstrate the consistent and significant efficacy of MissTime under different missing ratios across diverse downstream tasks. The code, datasets and implementation details are available at: https://anonymous.4open.science/r/MissTime-F956.