Alina Babenko, acting DS manager at Avito, published a detailed article on Habr titled "From CTR to Deals: How Avito's Monetization ML Models Work." For the first time, the production architecture of monetization ranking for one of Russia's largest marketplaces has been documented: a three-component system consisting of CTR, CVR, and a bid adjustment model with a custom loss function and probability calibration.
What happened
Alina Babenko, acting DS manager at Avito, published an article on Habr (AvitoTech Blog) on August 11, 2026, detailing the ML infrastructure for search and recommendation monetization. The system uses three models: the CTR model predicts the probability of a click, the CVR model estimates the probability of a target action after a click, and the bid adjustment model adapts the auction weight to the traffic's conversion rate. For training the adjustment model, query cross-entropy loss is used — a combination of standard log loss and group log loss, allowing budget redistribution among users without increasing the average seller CPC. Ranking quality is measured using stratified ROC-AUC, which is calculated within each query and aggregated by click volume. Probability calibration is controlled by the RIG = 1 − log loss / entropy metric. To combat position bias, a three-stage decorrelation using CatBoost with Inverse Propensity Weighting is applied. A plan to transition from CatBoost to deep learning with a focus on Airbnb's architecture is also announced.
Context
The AvitoTech publication is one of the rare cases of a transparent description of a production ML monetization pipeline for a major marketplace. In the advertising platform and marketplace industry, search and recommendation monetization issues usually remain closed. The documented approaches — CTR/CVR models, position bias mitigation, CatBoost for ranking — are widely known in themselves, but the specific combination of architectural decisions, metrics, and loss functions at Avito's scale had not been described before. The stratified ROC-AUC metric instead of global AUC is especially important for marketplaces with a long tail of queries, where global AUC masks quality on rare queries. RIG calibration is critical for financial calculations: uncalibrated probabilities lead to systematic errors in revenue forecasting. Query cross-entropy loss is a practical solution to a constrained optimization problem, which is usually solved using Lagrange multipliers.
Why this matters for the industry
The article provides a ready-made reference architecture for search monetization on marketplaces. Approaches to query cross-entropy loss for bid adjustment with an average CPC constraint, stratified ROC-AUC, and RIG calibration will be useful for DS teams of any advertising and marketplace platforms. The announced plan to transition from CatBoost to deep learning could become a benchmark for other Russian marketplaces: a wave of migrations from GBRT to DNN ranking is expected. Integrating dense embeddings with existing CatBoost is a realistic intermediate step. The publication will create demand for standard tools: ranking benchmarks, probability calibration frameworks, and multi-touch attribution platforms.
Why this matters for users
The material shows how problems that are usually only encountered in academic articles are solved in practice: long funnel attribution, probability calibration for financial calculations, bias mitigation in search results. It is useful for data scientists, ML engineers, and tech leads working with recommendation systems, search ranking, or AdTech. DS teams can adapt stratified ROC-AUC and RIG calibration for their systems right now. Teams that currently use global AUC and face evaluation bias will find immediately implementable alternatives. Startups can use the described methods as a baseline for their own ranking systems.
What is still unknown / limitations
The article only states a "planned transition" to deep learning, without architectural details, A/B test data, or benchmark comparisons. The focus on Airbnb's architecture is a research direction, not an achieved result. Inverse Propensity Weighting is statistically correct but sensitive to variance with small propensity scores, which may limit its effectiveness in practice.
Sources
- From CTR to Deals: How Avito's Monetization ML Models Work / Habr
- Monetization models — Alina Babenko, Avito / YouTube
Author
Look at AI, editorial team