Skip to content

Search, Ranking & Ads

Two case studies on ranking when the label is hard to trust. Search ranking centers on the relevance label problem: clicks are not relevance, and learning to rank has to cope with that gap. Ads ranking adds calibration, because a predicted click probability feeds directly into an auction, and adds the bias and delay problems that come with conversion labels.

Both sit one step beyond the recommendation skeleton: the model is similar, but the label and the layer on top, interleaving for search and the auction for ads, are where the real difficulty lives.