Feature Selection With GenSVM Using Group Lasso Regularization
An extension is proposed to the GenSVM classification algorithm by replacing the square penalty term by a group lasso regularization. A majorization for this new penalty term is derived and the extended model is implemented in Python. This new technique is then tested on two data sets and compared with the regular GenSVM. The group lasso extension is shown to behave as a feature selector, setting row vectors, corresponding to a specific attribute, of the SVM coefficient matrix collectively to zero. In terms of performance, the group lasso GenSVM is shown to be competitive to the base model.