I have my feature space that is

Year Month Day Hour Minute Second Lane Direction Speed Flag

A sample vector would look like:

[2015, 11, 5, 2, 20, 58, 4, 2, 45.7, 2]

The features take the values:

Year : 2012 - 2017

Month : 1 - 12

Day : 1 - 7

Hour - 0 - 23

Minute : 0 - 59

Second : 0 -59

Lane : 1 - 6

Direction : 1 - 2

Speed : continuous values

Flag : 1-3

My data is time series data. I am unsure of how to normalise each of these features. For example the year feature, is it necessary to normalise them? Most of my features take discrete values. If I do normalise them, would my SVM model perform better than when I don't normalise them?

Any suggestions on how to normalise time series data for SVM or reference to any papers will be appreciated.

  • 1
    $\begingroup$ does this question answer it? stats.stackexchange.com/questions/57010/… , I wouldn't really have anything to add beyond what is in there. $\endgroup$ – ReneBt Feb 21 '18 at 9:51
  • $\begingroup$ Also note that it is slightly weird to code time series like that. In most timeseries problems (and in most of papers on the topic), you encode time as a single continuous variable (e.g. the number of seconds is Unix Epoch) and set $t = 0$ for your first observation. SVMs also feel slightly nonstandard model for time series (vs. ARIMA, Gaussian processes). Not sure what your task is, but you may want to do some feature engineering to have binary features akin to "isWeekend". $\endgroup$ – Martin Modrák Feb 26 '18 at 15:06

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.