Diffusion Modeling and Transformer Knowledge Distillation Techniques for Time Series Imputation

In recent years, big data analysis has gained popularity, but data are often missing during collection, storage, etc., which affects the effectiveness of data analysis. Therefore, this paper proposes a Time Series Imputation method based on Diffusion model and Knowledge Distillation (TSI-DIKD). The diffusion model imputation missing values in time series and is combined with a Transformer structure to capture the intricate dependencies present in the time series data. Knowledge distillation is also used in the reverse process of the diffusion model to improve imputation accuracy. We conducted experiments on healthcare and environmental data to compare the performance of the model with different percentages of missing values. The results show that our model outperforms other models on all three evaluation metrics.