|
What is the state of the art for ordinal regression? Is it better to treat it as regression or as multiclass? I have labels 1-10 of integers and 1,5,8,9 are easily the most populated, with 10 being rare. Can someone point out good ordinal regression paper? Thanks |
|
You should look at the paper Gaussian processes for ordinal regression, and another paper that uses the margin based approach New approaches to support vector ordinal regression. AFAIK, both are kind of state-of-the-art algorithms; though the GP one came out a bit later and has shown better performance as compared to the margin based approach. Of course, GPs can be slow but there has been lots of work to speed up GPs. 1
Just to add, there is code available for both the above papers: The GP ordinal regression code: http://www.gatsby.ucl.ac.uk/~chuwei/code/gpor.tar Support vector ordinal regression code: http://www.gatsby.ucl.ac.uk/~chuwei/svor.htm
(Jul 26 '10 at 15:17)
spinxl39
|
|
This paper uses large margin methods for ordinal regression. This paper explores a learning to rank approach that is based on ordinal regression. I've seen a lot of papers (and some tutorials) on learning to rank, but I haven't seen much in ordinal regression. I guess a classification approach with assymmetric losses (ie, classifying a 1 as a 3 is worse than as a 2) might be a good start. |
