I have sets of features of different nature(for example, 300 features from FFT-transform, 1000 categorical features and so on). However there are only 900 samples and I`m trying to select important features using Lasso. So the question is: should I perform feature selection differentially on subsets of features or better firstly, concatenate features and only then perform feature selection?
Let LASSO pick the best ones. If the features are highly correlated and you want them picked as a group, add some L2 regularization too. This is called Elastic Net regularization, and it is a generalization of L1 and L2 regularization. Other than that, do not feel obliged to artificially group features.