Of the two best known techniques for feature scaling in Machine Learning:

  • Normalizing a feature to a $[0, 1]$ range, through $x - x_{min} \over x_{max} - x_{min}$


  • Standardizing the feature (also referred to as z-score), through $x - μ \over σ$, where $μ$ is the mean and $σ$ is the standard deviation.

Is there any reason to prefer one over the other? Does any one outperform the other when used with certain algorithms?

  • 1
    $\begingroup$ It depends on the objective. $\endgroup$ – user2974951 Sep 14 '18 at 12:32
  • $\begingroup$ @user2974951 Is there any case where one is preferred over the other? Or does one have any useful properties compared to the other? $\endgroup$ – kfn95 Sep 14 '18 at 12:35
  • $\begingroup$ Your comment makes me believe this is a possible duplicate of Is it a good practice to always scale/normalize data for machine learning? $\endgroup$ – Frans Rodenburg Sep 14 '18 at 12:53
  • 1
    $\begingroup$ @FransRodenburg The author asks which feature scaling to perform, while the question you referenced is about whether or not to use feature scaling. I think it's a valid question. $\endgroup$ – user204007 Sep 14 '18 at 14:17
  • $\begingroup$ @JohnDoe I realise the difference in titles, but if you read the accepted answer you'll see that it is more similar than it might seem at first. $\endgroup$ – Frans Rodenburg Sep 14 '18 at 14:19