Unbiased Learning to Rank
Click logs look like free training data — millions of signals about what people wanted. They are not. A result gets clicked partly because it sat at the top of the list, and an item the old ranker never showed collects no clicks at all. Unbiased Learning to Rank trains on logged clicks while correcting for the fact that a previous policy decided what the user could ever see.
Correcting the click signal
A click counts as evidence only in proportion to how likely that result was to be examined. Each training example is weighted by the inverse of its examination propensity — inverse propensity scoring applied to a ranking loss — so a click at rank ten carries far more weight than a click at rank one. The propensities are estimated, not assumed: by swapping neighbouring results on a small slice of traffic, by harvesting cases where the same item appeared at different ranks, or by fitting a click model with regression-EM over existing logs.
What it prevents
- Training naively on clicks teaches a ranker to reproduce the ranking that produced them, popularity bias included.
- Relevant items the previous policy buried are read as negatives, so they stay buried — a loop that more data does not break.
- It makes a live system's own logs usable without paying for human relevance judgements.
Where it sits
The correction belongs to the ranking stage, where a LambdaMART or neural ranker orders a few hundred candidates, and it is the training-time twin of counterfactual evaluation: the same propensities that debias the loss also debias the offline metric. Without it, offline gains keep failing to reproduce online.
Unbiased Learning to Rank trains a ranking model on logged user clicks while correcting for the biases in how those clicks were collected — chiefly position bias, where higher-ranked results are examined more, and the selection bias of items the previous ranker never showed. Training examples are reweighted by an estimated examination propensity, usually through inverse propensity scoring. Without it, a ranker learns to imitate the policy that generated its training data instead of improving on it.
Imagine a shop where everything at eye level sells well. You could conclude that customers love those products, but really they just saw them first. The things on the bottom shelf never had a chance.
Recommendation software has the same problem. It learns from what you clicked, and you mostly click whatever it already put at the top. Unbiased Learning to Rank is the fix. Before learning anything from a click, the system works out how much of that click was genuine interest and how much was simply good placement, then gives extra credit to clicks on things that were hard to notice. Without that correction, the software just keeps recommending what it already recommended.