AI for feature selection in the context of geomagnetic solar events prediction

In statistical learning and, specifically, in supervised learning enhancing sparsity is a crucial issue. In the machine learning (ML) framework, feature reduction techniques allow to identity which dimensions of a given dataset are more relevant to a specific task, then supervised learning models are trained with that reduced number of significant features. The first step is usually performed by means of linear models such as Lasso regression (or variations of it), Recursive Feature Elimination (RFE) and linear Support Vector Machines (SVMs), and once the features are identified non-linear models are exploited to predict the given task. Here we introduce a new approach, called in what follows greedy feature selection, which allows to perform an iterative selection depending on the classifier that has been chosen for the prediction phase. This technique is specifically applied in the realm of solar physics for the prediction of geo-effective manifestations of the active Sun, to understand which physical parameters are decisive for classifying an event. The models that have been considered include non-linear SVM and a Feed-forward Neural Network (FNN). Regarding the prediction phase, the selection of features with non-linear greedy method is then compared with state-of-art ML algorithms. Numerical results show that, usually, when only the few greedily selected relevant features are considered, classification performances are improved.