standardization vs normalization

Both are feature scaling and preprocessing steps.

• standardization does not change the distribution!
  • column-based
  • standard scaler (μ=0, σ=1)
  • min max scaler (min=0, max=1)
• normalization changes the distribution, and has two-types. L1 and L2
  • row-based
  • L1 : least absolute deviations, sum of absolute rows = 1
  • L2 : least squares (root), sum of squares rows = 1

                                   scaling     
       gradient calculations  |   mandatory      | linear r, logistic r., neural n., deep l.
 distance based calculations  |    no-need       | tree, ensemble models, k-means ...
 
         neural networks, dl  |  Normalization   | cause model expects normalized values!
    linear, logistic r., svm  |  Standardization