Towards Synthetic Multivariate Time Series Generation for Flare Forecasting

One of the limiting factors in training data-driven, rare-event prediction algorithms is scarcity events interest resulting an extreme imbalance data. There have been many methods introduced literature for overcoming this issue; simple data manipulation through undersampling and oversampling, utilizing cost-sensitive learning algorithms, or by generating synthetic points following distribution existing While generation has recently received a great deal attention, there are real challenges involved doing so high-dimensional such as multivariate time series. In study, we explore usefulness conditional generative adversarial network (CGAN) means to perform data-informed oversampling order balance large dataset We utilize flare forecasting benchmark dataset, named SWAN-SF, design two verification both quantitatively qualitatively evaluate similarity between generated minority ground-truth samples. further assess quality samples classical, supervised machine algorithm on data, testing trained model unseen, The results show that classifier augmented with series achieves significant improvement compared case where no augmentation used. popular evaluation metrics, TSS HSS, report 20-fold 5-fold improvements, respectively, indicating remarkable statistical similarities, CGAN-based complicated tasks forecasting.