TIformer: A Transformer-Based Framework for Time-Series Forecasting with Missing Data

Zuocheng Ding, Yufan Chen, Hanchen Wang, Xiaoyang Wang, Wenjie Zhang, Ying Zhang

Abstract: Long Sequence Time-series Forecasting (LSTF) is a fundamental problem with wide real-world applications. However, missing data in time series is ubiquitous, posing significant challenges and affecting the accuracy of long sequence time-series forecasting. To address this issue, we propose a comprehensive framework Time-series Imputation transformer, namely TIformer, for imputing missing data and subsequently conducting time series forecasting. There are two major modules of TIformer, the missing data imputation module and the time series forecasting module. In the first module, we employ data imputation methods to improve the data quality for downstream forecasting. In the second stage, we employ the transformer-based model for long-term time series prediction, which leverages the frequency information on time series. TIformer is an effective work on improving time series forecasting performance on scarce data. Through experiments on various datasets, we have demonstrated the powerful effectiveness of our method in improving prediction accuracy.