MinMaxScaler vs StandarScaler

I have some doubts on Normalisation (or Standardisation, whichever term you prefer).
Any inputs will be really helpful:

  • While standardising/normalising a data set, how do I decide whether to do min max scaling or do standard scaling (0 mean, unit variance)?

  • Normally we consider normalisation at the level of data instead of at the level of individual feature. For e.g. we pass all of X_train to MinMaxScaler or StandardScaler (sklearn). Is it possible (and common) that data may have several features and some of those features follow normal distribution while other follow uniform? What’s the best approach for such cases

  • How do I check distribution of data with many features (lets say > 50)? Do I need to check distribution of every feature (by plotting for example)

Right now, the approach I follow is to just try both on all of data, and whichever gives better result, accept it. What are other better popular approaches?


Hi @sanjayk,
The first suggestion would be to refer the shared links in How and when to apply standardization and normalization in the dataset..

Many a times, data is normalized with some other methods, i.e Batch Norm, Instance Norm, Layer Norm etc.

1 Like