Long Short-Term Imputer: Handling Consecutive Missing Values in Time Series

Jiacheng You, Xinyang Chen, Yu Sun, Weili Guan, Liqiang Nie

Abstract: Encountered frequently in time series data, missing values can significantly impede time-series analysis. With the progression of deep learning, advanced imputation models delve into the temporal dependencies inherent in time series data, showcasing remarkable performance. This positions them as intuitive selections for time series imputation tasks which assume ``Miss Completely at Random''. Nonetheless, long-interval consecutive missing values may obstruct the model's ability to grasp long-term temporal dependencies, consequently hampering the efficacy of imputation performance. To tackle this challenge, we propose Long Short-term Imputer (LSTI) to impute consecutive missing values with different length of intervals. Long-term Imputer is designed using the idea of bi-directional autoregression. A forward prediction model and a backward prediction model are trained with a consistency regularization, which is designed to capture long-time dependency and can adapt to long-interval consecutive missing values. Short-term Imputer is designed to capture short-time dependency and can impute the short-interval consecutive missing values effectively. A meta-weighting network is then proposed to take advantage of the strengths of two imputers. As a result, LSTI can impute consecutive missing values with different intervals effectively. Experiments demonstrate that our approach, on average, reduces the error by 57.4% compared to state-of-the-art deep models across five datasets.

Submission Length: Regular submission (no more than 12 pages of main content)

Changes Since Last Submission:

First, in the previous version, we followed the dataset splitting strategy from the original TimeFlow paper, which led to suboptimal experimental results. In the camera-ready version, we carefully tuned the TimeFlow in experiments, resulting in a substantial improvement in performance, though it required significantly more time and memory. The updated results can be found in Tables 15 and 16. Overall, under the Blackout missing pattern, TimeFlow incurs a computational cost roughly 4-5 times higher than LSTI, yet LSTI continues to outperform TimeFlow.

Second, we added Table 10 to provide detailed information about the newly added datasets in the appendix.

Assigned Action Editor: ~Jacek_Cyranka1

Submission Number: 3706