Improving Flare Detection via Masked Difference Prediction

Recent years have observed the rapid development of astronomy observation devices, hence leveraging a large amount of observation data to automatically detect flare has become an emerging research topic. Previous studies on the flare detection task focus on using hand-drafted astronomy features or time-series analysis to capture the abnormal values in the luminosity data. However, these approaches heavily rely on domain expertise and are difficult to transfer into other stars or special phenomena. In this paper, we consider adopting deep learning technology into this task. To enhance the transferability and build an effective model, we propose a novel task, namely a masked difference prediction task to learn the enhanced representations of each luminosity difference and the whole sequence. The learned representations can be transferred into conventional RNN and CNN models with simply fine-tuning on the original flare detection task. Experiments show that our approach can bring improvement to CNN and RNN models.