I'm new to machine learning and try to clarify my problem in research. I just wonder if I can compare the importance of two different variables in two different sorts. For example, A and B are two variables that I want to compare their contribution to ML accuracy. The rest of the variables (like C, D, and E) for each sort are the same.

If I got the results that the rest of the variables (C, D, and E) contributes more (obtain a higher importance score) in the model predicted by A and the rest, can I just say B performs better for ML and its contribution is more significant than the model with A and the rest?

  • $\begingroup$ What kind of model do you plan to use? If you need to measure variable importance I recommend linear regression or decision trees. $\endgroup$
    – bstrain
    Commented Dec 17, 2019 at 9:29
  • $\begingroup$ Actually I have already got the variable importance list. Just wondering if I can compare the importance of two variables in two different sorts. Which one is more significant to accuracy. $\endgroup$
    – Ang Ji
    Commented Dec 18, 2019 at 10:32

1 Answer 1


No you can not. You are forgeting about variable interactions (C-A,D-A,E-A, etc...) that could favor A.

You could answer following qestion: If I were to measure information in my variables via variability how would I proceede? For example PCA

The more information you have in the variable(predictor) the higher the chance that it will contribute more to accuracy

  • $\begingroup$ OK, I'll check this later. So does it mean that I can just use statistic methods to identify which one contributes more to accuracy, instead of machine learning? $\endgroup$
    – Ang Ji
    Commented Dec 18, 2019 at 10:35
  • $\begingroup$ a) whats the difference b) Yes, you can just use statistic methods but not only. I gave you an example of how you can make reasonable comparisons c) Accept if this answer satisfies you $\endgroup$
    – Noah Weber
    Commented Dec 18, 2019 at 10:38
  • $\begingroup$ Correct. Those comparisons only hold when the models don’t change $\endgroup$
    – HEITZ
    Commented Dec 19, 2019 at 7:06

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.