|
The last decade has seen an increasing body of work in machine-learned ranking. Is there any consensus on what is the state of the art in this area, if not at the level of specific algorithms, at least in terms of the general approach (pointwise, pairwise or listwise) ? I'm mostly interested in ranking applied to contextual advertising and sponsored search (CPC and/or CPA pricing). |
|
Looking at the winners of the recent yahoo learning to rank challenge (which used data and had submissions from people working on state-of-the-art learning-based search engines) it seems that a listwise approach is favored (as both the true loss and the loss used as a proxy in the challenge are listwise), and boosted regression trees are the learning method of choice. The challenge unfortunately didn't touch on contextual advertising or sponsored search, and also in these areas I think the state of the art is probably hidden inside the companies working in these areas. Looking at some related papers I get the impression that these tasks are modeled with reinforcement learning, but this is completely unfounded. Not completely unfounded... as you say, a lot of papers coming out of Yahoo, Microsoft, Google, etc. are about decision making (aka reinforcement learning, aka bandit problems). Look at John Langford's work for example. A very relevant paper is Non-stochasitic bandit slate problems
(Mar 01 '11 at 05:07)
Noel Welsh
|
