1
$\begingroup$

I have a sensor dataset with 16 features as numerical values (12 are continuous and 4 are discrete). I am using LSTM model to fit the data and do some classification. As both continuous and discrete are numeric values, is it fine if I normalize both continuous and discrete instead of separating them and only normalizing continuous ones. I saw one way is that to separate them and only normalize continuous values, and then combine them together again.

Most answers suggest to separate the data if the discrete values have some ordinal and nominal data, and then further process them by using one-hot encoding or continuousification. However, in my case all the data are numeric. Therefore, I think these solutions might not be applicable to my case.

I did both (normalize both together) and (only normalize continuous and then combine them again). I did not see much difference in terms of the final results. However, I am not sure if this is theoretically correct or not. I do not want to do something might be wrong so I am looking for a solid theoretical answer.

$\endgroup$

1 Answer 1

1
$\begingroup$

The reason for normalizing is to make optimization easier, so yes it makes sense in the situation you described to normalize discrete values as well.

$\endgroup$
3
  • $\begingroup$ Thanks @MONODA43. Yes, I agree with. It makes sense. But, I was worried if this might be wrong and might not be applicable to other situations so I want to know more about it. $\endgroup$
    – Amhs_11
    Sep 25, 2020 at 2:57
  • 1
    $\begingroup$ I think you've already considered the other possibilities such as learning an embedding for these features. Since they are numeric it probably makes more sense to do as you've done. $\endgroup$
    – dmh
    Sep 25, 2020 at 3:39
  • $\begingroup$ Cool, thank you very much @MONODA43. $\endgroup$
    – Amhs_11
    Sep 25, 2020 at 3:41

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.