Improving Group Lasso for high-dimensional categorical data