Why is training a ranker on observed clicks without considering display position problematic?
Classic MLLevel: IntermediateXGBoost · Learning to rank
Why is training a ranker on observed clicks without considering display position problematic?
Hints
- A result cannot be clicked without exposure.
- The current ranking policy affects labels.
Answer
Exposure at higher ranks increases click opportunity, so labels mix relevance with presentation policy. Position-debiasing methods require assumptions and evaluation; raw clicks are not unbiased relevance judgments.
Source and licence: XGBoost · Learning to rank · Apache-2.0 · answer: Quantoros editorial