arXiv2022
Recently, there has been growing interest in the use of machine-learning methods for predicting solar flares. Initial efforts along these lines employed comparatively simple models, correlating features extracted from observations of sunspot active regions with known instances of flaring. Typically, these models have used physics-inspired features that have been carefully chosen by experts in order to capture the salient features of such magnetic field structures. Over time, the sophistication and complexity of the models involved has grown. However, there has been little evolution in the choice of feature sets, nor any systematic study of whether the additional model complexity is truly useful. Our goal is to address these issues. To that end, we compare the relative prediction performance of machine-learning-based, flare-forecasting models with varying degrees of complexity. We also revisit the feature set design, using topological data analysis to extract shape-based features from magnetic field images of the active regions. Using hyperparameter training for fair comparison of different machine-learning models across different feature sets, we show that simpler models with fewer free parameters \textit{generally perform better than more-complicated models}, ie., powerful machinery does not necessarily guarantee better prediction performance. Secondly, we find that \textit{abstract, shape-based features contain just as much useful information}, for the purposes of flare prediction, as the set of hand-crafted features developed by the solar-physics community over the years. Finally, we study the effects of dimensionality reduction, using principal component analysis, to show that streamlined feature sets, overall, perform just as well as the corresponding full-dimensional versions.