Feature binarization refers to the mapping of a continuously valued feature to a binary one. It is essentially equivalent to the thresholding strategy used in binary classification, but applied to features instead. It can be viewed as a special case of scalar quantization where only two labels are used.