3
$\begingroup$

If I'm using a prediction model which uses a distance-based metric to help calculate class separability, why would I use Min/Max over the Z-Score? I've always thought Z-Score as superior because during your training stage you can't state if you've seen the minimum and maximum values, meaning during production there may be data samples that get normalized to a range not seen by the prediction model, whereas if you use the standard deviation and mean to calculate a Z-Score you can minimize this issue.

Does anyone have any reasons for why Min/Max scaling can be better suited rather than using the Z-Score?

$\endgroup$

1 Answer 1

2
$\begingroup$

There are ML algorithm which react sensible to the change of distance ratio betw. two samples, think k-NN with euclidean distance function. With k-NN a min-max scaling pre-processing could be a better alternative.

But as always, it depends highly on your data set, how many outliers you have, etc.

And for production, i would argue that in both cases you don't exactly know what comes your way. the only think to remember here, is that you scale the new data instance the same way you scaled the other values.

Then it should not make a difference for your ML algorithm.

IMO, a good walk-through on pros/cons is the Scikit-learn User Guide.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.