tag:blogger.com,1999:blog-6569681.post890829005069744071..comments2024-01-15T13:17:33.771-08:00Comments on Geeking with Greg: LambdaRank, RankNet, and MSN SearchGreg Lindenhttp://www.blogger.com/profile/09216403000599463072noreply@blogger.comBlogger2125tag:blogger.com,1999:blog-6569681.post-87184310943282938332008-01-15T15:03:00.000-08:002008-01-15T15:03:00.000-08:00Hi, Jon. Sorry, I was not clear in my post. I me...Hi, Jon. Sorry, I was not clear in my post. I meant "underweight the value of the first search result and overweight the value of later results" only in comparison to other possible definitions of DCG.<BR/><BR/>I'll make a correction. Thanks for pointing out the issue.Greg Lindenhttps://www.blogger.com/profile/09216403000599463072noreply@blogger.comtag:blogger.com,1999:blog-6569681.post-31278967382688375442008-01-15T14:53:00.000-08:002008-01-15T14:53:00.000-08:00I believe your interpretation of NDCG is not quite...I believe your interpretation of NDCG is not quite right. The "discount" in NDCG penalizes the score for relevant documents further down the list. In this paper, this discount is: c=1/log_2(1+i), where i is the rank of the document. So, the document at the top of the ranked list (i=1) would receive a discount of 1.0, at rank i=2 a discount <1.0, etc.Anonymousnoreply@blogger.com