I have a dataset with 5K records focused on binary classification problem. I have about 60 features.

Out of 60 features, around 45-46 features are of 'Min' and 'Max' type.

For example, minimum blood pressure, maximum old pressure, minimum heart rate, maximum heart rate, minimum potassium, maximum potassium etc.

Several other vital parameters like sodium, urea etc. follow this pattern of min and max.

Any suggestions on how can I transform them without losing info . currently when I use as-is, I get only around 86% accuracy with recall value of minority class only around 70. Resamplin didn't help much.

Your suggestions and experience in how to transform this to yield better predictions would really be helpful.


1 Answer 1


ML has sometimes problems to find relations obvious for people, so I would help a little bit by creating a new feature:

relX = (maxX-minX)/minX

because some diseases are more correlated with jumps of some parameters than with their level, however, such relX may be too much correlated with target and may cause overfitting. Usually, after creating such a feature, it is better to remove one of the original features, sometimes both.

  • $\begingroup$ Hi, thanks for the response. Upvoted.may I know what does RelX stand for? I mean am trying to understand how did you arrive at this formula?. $\endgroup$
    – The Great
    Dec 16, 2019 at 8:14
  • $\begingroup$ Am searching online too.. But couldn't find anything. Is there any name for this formula or equation? $\endgroup$
    – The Great
    Dec 16, 2019 at 8:27
  • $\begingroup$ MinX and maxX are for example minimum and maximum of blood pressure for a given patient. Difference of these features in relation to minX, maxX or (minX+maxX) may be very valuable, especially, that in the case of deep learning - after normalization of these raw features regarding all patients, this relation will be hard to reconstruct. Is the name important? Let's call it a relative difference of extremal values of given parameters. $\endgroup$ Dec 16, 2019 at 8:43
  • $\begingroup$ Sure thank you. So I should construct these features on my raw original raw data data (non-standardized) data. Once I create this new feature for each measurement(bp,hr,temp etc), I can standardize them. Is my understanding right? $\endgroup$
    – The Great
    Dec 16, 2019 at 9:15
  • $\begingroup$ Exactly. Good luck. $\endgroup$ Dec 16, 2019 at 9:17

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.