Using LASSO regularization as a feature selection tool.

Abstract: The subject of deep learning has become increasingly popular, especially for machine learning applications where a large number of input variables have to be processed. However, there are instances of problem solving, where a full understanding of the variables is of high importance. When dealing with data sets containing a large number of input variables, the established methods of feature selection require a considerable time investment. Regularization is a method typically associated with prevention of overtraining, but in this study, the possibility is explored of using LASSO regularization as a feature selection tool. The input variables of several data sets were ranked with respect to a measure of synaptic weight magnitude. A conclusion was drawn that this method is a very fast and efficient way of filtering out less important variables.