I have binary columns in my dataset (20) e.g. hot_weather, discount (y or no), where in each case 1 = yes no = 0. I am using this data on tree based methods.

It is a regression problem and my RMSE is around 1500! Running feature importance reports I get 0 for all but 2 of these binary columns which makes me think they cannot be good for my tree based model.

How else can I transform these binary columns in my data?

  • 1
    $\begingroup$ First some context questions: RMSE=1500 may be excellent or terrible, depending on the scale of your target; what is the scale of yours? Are there continuous variables in your model as well? How does a simpler model do? And now, a non-answer opinion: it will be very hard to give good advice for feature engineering without the actual full information about your problem: what is the target, what are these binary features, etc. It may well be that there is nothing better to find. $\endgroup$
    – Ben Reiniger
    Feb 6, 2021 at 21:31
  • $\begingroup$ the scale of my target is no of products sold. What simpler model could i use? linear regression? I have not actually normalized my data , 10 of the features are continuous the other 20 are binary. examples of continuous are temp, no of clicks on website for that day, temp, etc.. some of the data is bimodal dsitributed others skewed, otherwise binary. This is why i opted to use tree based algorithms as it is invariant to underlying distrubution. I also thought that if i normalized the data it would become difficilt to explain to stakeholders.. @Ben Reiniger $\endgroup$
    – Maths12
    Feb 7, 2021 at 11:44


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.