0

Wikipedia states three methods of feature scaling. Which should be used when? (what are the considerations?)

To be specific, I need it for Sentiment Analysis on phrases, implemented with SVM.

(NOTE: I've seen this post. It explains the different methods quite well, but doesn't say anything about when each should be used).

Thank you :)

Community
  • 1
  • 1
Cheshie
  • 2,777
  • 6
  • 32
  • 51
  • Which application you are targeting is by and large irrelevant. The answer depends entirely on what your feature model looks like. – tripleee Jan 16 '14 at 09:14

1 Answers1

0

Actually this is quite hard to give any reasonable rules for selecting scaling over standarization. Standarization of your data has a good theoretical justification and is less influenced by outliers than scaling. As the result the most commonly used method of preprocessing is standarization.

In particular, if you ask about standarization than you use some kind of bag of words representation of your data. In such a case tf-idf is the most obvious selection of the data representation, which in fact is hardly influenced by any scaling/standarization as it is quite well standarized itself (due to the internal normalization and log-scaling).

lejlot
  • 64,777
  • 8
  • 131
  • 164