Using multiple instance learning for explainable solar flare prediction

In this work we leverage a weakly-labeled dataset of spectral data from NASAs IRIS satellite for the prediction solar flares using Multiple Instance Learning (MIL) paradigm. While standard supervised learning models expect label every instance, MIL relaxes and only considers bags instances to be labeled. This is ideally suited flare with that consists time series UV spectra measured along instrument slit. particular, consider readout window around Mg II h&k lines encodes information on dynamics chromosphere. Our are not able predict whether occur within next $\sim$25 minutes accuracies 90%, but also explain which profiles were particularly important their bag-level prediction. can used highlight regions interest in ongoing observations real-time identify candidates typical precursor profiles. We use k-means clustering extract groups appear relevant The recovered show high intensity, triplet red wing emission single-peaked h k lines, as found by previous works. They seem related small-scale explosive events have been reported tens before flare.