IBMA: An Imputation-Based Mixup Augmentation Using Self-Supervised Learning for Time Series Data

The introduction of Imputation-Based Mixup Augmentation (IBMA) marks a significant advancement in time series forecasting, a field that traditionally lacks robust data augmentation techniques compared to others like image processing. By integrating imputation methods with Mixup augmentation, IBMA enhances model generalization and forecasting performance. Evaluated across several state-of-the-art models, including DLinear, TimesNet, and iTrainformer, IBMA demonstrated its effectiveness in experiments on four datasets (ETTh1, ETTh2, ETTm1, ETTm2), achieving 22 performance improvements out of 24 instances. This includes 10 instances where it delivered the best results, particularly with the iTrainformer model. The consistent enhancements provided by IBMA highlight its potential to significantly improve predictive accuracy in time series analysis, addressing a critical gap in the field and paving the way for more sophisticated forecasting methodologies.