A Study of Feature Selection of Magnetogram Complexity Features in an Imbalanced Solar Flare Prediction Data-set

Solar flares are the conversion of stored magnetic energy into particle acceleration and radiation, with potential significant detrimental effects on earth including damage to technological infrastructure. Recent work has considered methods to predict flare activity from quantitative measures of the solar magnetic field. Feature selection methods provide insight into measures which have the largest discriminatory potential and provide a means to streamline real-time processing of solar data. Since solar flares are rare events, data-sets for such predictive analysis are inherently imbalanced, causing a bias in the classification. Monte Carlo experiments with randomly sub-sampled, balanced data-sets mitigate classifier biases for imbalanced data-sets, but it is unclear how to implement and interpret feature selection in such a framework. We propose a method to determine a feature subset within a sub-sampled classification based on a histogram analysis of selected features. We show that the feature subsets resulting from this analysis yield better classification accuracies across a large imbalanced data-set unseen in the feature selection and classifier training.