The Group-Lasso for generalized linear models

摘要: The Group-Lasso method for finding important explanatory factors suffers from the potential non-uniqueness of solutions and also high computational costs. We formulate conditions uniqueness which lead to an easily implementable test procedure that allows us identify all potentially active groups. These results are used derive efficient algorithm can deal with input dimensions in millions approximate solution path efficiently. derived methods applied large-scale learning problems where they exhibit excellent performance testing helps avoid misinterpretations solutions.