I have a raw unlabeled dataset, and I want to design a model to perform a regression. In my dataset, it does not make sense to give each observation a value, but it does make sense to sort them. Can I implement an algorithm to create values for each observation by sorting them?

I thought about this:

  1. Select N random observations and sort them
  2. Give each observation a new score, equal to its position
  3. Calculate the score of an observation as the average position across all times the observation was picked
  4. return to step 1

Does it make sense? Is there any machine learning branch that studies this kind of scenarios?


1 Answer 1


There is a field for this called ordinal regression. Does each unique observation have it's own rank or can observations share a rank? IE. if you have 10 elements, are they labeled 1,2,3 ... 10? or could it be 1,1,1, 2,2,2,3,3,3,4? can't post as comment due to karma What are these values supposed to represent? Why are you doing this analysis?

  • $\begingroup$ Many thanks for your help. Each observation could have its own rank or share a rank. It is an arbitrary decision. I guess it is simpler to suppose each one have its own rank $\endgroup$ Oct 5 at 6:22
  • $\begingroup$ So are you trying to create a model to pit item A vs item B and decide which is higher ranked? Using the attributes to explain why item A or B is higher ranked? $\endgroup$
    – user70889
    Oct 5 at 13:08
  • $\begingroup$ Yes to the first question. No to the second: I don't need an explanation of why A>B. $\endgroup$ Oct 11 at 10:38
  • $\begingroup$ Will you always be comparing just two elements at a time? $\endgroup$
    – user70889
    Oct 11 at 13:08
  • $\begingroup$ No, I thought about comparing in batches of 10 elements (N=10). $\endgroup$ Oct 12 at 15:56

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.