I want to obtain a total ranking from pairwise binary comparisons. For this, I can use algorithms like Balanced Rank Estimation or Bradley-Terry Model. However, I wonder if you need fewer comparisons, if the compared pairs are chosen cleverly. So rather than having a complete set of comparison results, we compute the next comparison for the participant on the fly.

The algorithm I'm looking for estimates a ranking that eventually converges. After each comparison, the estimate is improved and a meaningful new comparison pair is suggested, to help the algorithm to converge as fast as possible.

Which algorithm can I use for this task?


I found a talk of Wiki Surveys creator Matthew Salganik talking about my question. He refers to a formal way of defining what's the pair that likely will give you the most information. Unfortunately, I couldn't find an algorithm or paper on this. How are those kinds of statistical models called and how can I find out more about them?



Your Answer

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