Indeed, this question is all researcher's question which are working in Data Mining field, I think you could use the features which is important and they have main effect on data clustering or classification, as my experience, I propose using Rough Set feature selection algorithm aim this issue, here I want to share with you my experience :
Weighted PCA, PLS, SVM, RF (Random Forest) and Hybrid methods are used for feature selection and extraction.
All of this methods have their own advantages and neither of the existing methods may be preferable to others in all situations. In my view, RF is powerful and can be use for loathe feature extraction/selection and regression. This was provided in "R" which was successfully applied in above mentioned applications!