The Top of the List: Ranking, Learning to Rank, LambdaMART, Clicks and Ad Auctions

Deep End Era. Engineering with Sankalp · Oct 28, 2026 · 50m

Orchard's new click model beats the old one on every classification metric, and makes search worse: fewer purchases, more reformulated queries. Leila's team has built a model that is good at scoring when they needed one that is good at ordering. Part ten of The Deep End's Machine Learning series starts in a Cranfield college library, computes NDCG by ear to show why swapping the top two results costs fifteen points while swapping positions three and four costs one, and follows Microsoft's path from RankNet to LambdaRank, where a coach's attention goes to the swaps that would change the race…