Importance measures derived from random forests: characterisation and extension