Problem
Avito ranks ads in search partly by ECPI, the expected revenue per impression. ECPI multiplies the funnel probabilities (click, contact, deal) by the commission, so each model must output a calibrated probability, not just a ranking score. Calibration depends on how training labels are assigned. A contact often comes later than the impression that caused it, for example from favorites or a recommendation feed, and one impression can be followed by several contacts. The label "contact on this SERP = 1" loses such contacts, and simple last-touch fixes distort the number of positives.
Example from the article: five SERPs and three contacts. One contact falls between the third and fourth SERP and matches no row.
Solution in brief
Give every impression (SERP) that precedes a contact a positive label, and spread each contact over its own chain of earlier SERPs with weights from a beta distribution. Weights of one contact sum to 1, so the total weight equals the number of contacts. A final correction of positive weights makes the product of the stage models converge to deals per impression.
Questions this case answers
- How to assign training labels to earlier exposures when the positive outcome happens later and one exposure can be followed by several outcomes
- How to weight the first, middle and last touches of a chain of exposures with a prior belief about their importance while keeping predicted probabilities calibrated
- How to assign training labels to search impressions when a contact or deal happens later and not on the impression we showed
- How to get calibrated CR and deal-probability models for ranking by expected revenue
Tools and methods
- Weighted multi-touch attribution — each contact distributes weight over the SERPs before it.
- Beta-distribution position weights — the first and last touches get more weight than the middle, set by two priors.
- Weight correction for the model chain — makes equal to .
- ROC AUC, RIG, calibration — offline metrics for the comparison.
Google or a one-time email code. No password.