Abstract
Time series data are widely used in critical sectors such as finance, healthcare, and environment to analyze temporal trends and patterns for prediction, monitoring, and decision-making operations. However, these datasets often suffer from noisy labels, which can significantly degrade the accuracy and reliability of the analysis. Existing research tends to focus on noisy labels in regular time series data while overlooking the unique complexities presented by irregular time series (ITS) data. In ITS, the likelihood of noisy labels is higher than in regular data due to the obscure or complex patterns resulting from uneven observation intervals and missing data ratios, which contribute to more frequent labeling errors. This paper aims to address the noisy label problem in ITS data by designing a novel risk estimator for effective analysis. We carefully investigated the potential relationship between noisy patterns and data irregularity and used the findings to inform the estimation process. Our results show that our proposed method outperforms existing approaches.