Attention with Mixture Experts model for Multivariate Time Series Imputation

Multivariate time series data often suffer from missing data problems due to sensor failures and environmental impacts, which destroy the integrity of the time series and seriously affect the interpretability of the data. Insufficient long-term dependency capture and difficulty in capturing complex non-linear relationships between multiple features in existing imputation models. To address these issues, we propose an attention with mixture experts model (AME) for multivariate time series imputation. The model consists of two core modules: the diagonal attention based on time bias (DATB) module enhances temporal modeling capability by introducing time bias to capture temporal dependencies; and the mixture of experts (MoE) module selects appropriate expert combinations for different features through routing mechanisms to capture the complex non-linear relationships between features better. We conducted extensive experiments on four real-world datasets, and the results showed that the AME model significantly outperformed the existing five methods in imputation accuracy, validating the effectiveness of the proposed method.