0
$\begingroup$

I want to predict properties of a physical system (e.g. dimension, weigth, number of items) based on a set of known attributes for the system. I have a good, large database of executed systems together with a catalogue of many properties for the systems.

Yet I am not sure which attributes are best and orthogonal to each other, that's why I decided to try Partial LeastSquare Regression.

But the choice of attributes and data types is wide, I have discrete categories ("A", "B", …) , continous (float) variables within limits (e.g. ambient temperature), booleans, integers etc. What is the best practice to implement a regression with such a bucket of different independent variables? Map everything into normalized floats?

Actually, I have to use multiple mulitvariate regression as the dependent variable is a vector. It also consists of integers and floats. So the question relates also to the output result.

Improve this question
$\endgroup$
2
  • 1
    $\begingroup$ For discrete variables, you can always use one-hot-encoding. For booleans, you can use simply use 0 for false and 1 for true (or vice-versa). For integers, there is nothing to be done since they are usable as is. $\endgroup$
    – gunakkoc
    Jul 2, 2018 at 20:32
  • $\begingroup$ For the output, same logic applies. Discrete variables can again be coded to 1 and 0 for each level. However, when you want to predict a discrete variable, one way is to choose the column, which refers to a level, having highest value. See stats.stackexchange.com/questions/218125/… or you can search for one-hot-encoding. $\endgroup$
    – gunakkoc
    Jul 2, 2018 at 20:36

0

Reset to default

Your Answer

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