Winner's Solution Walkthrough: 1st Place (AvitoTech ML Cup 2026)

Andrey Miroshnikov, Anton Legchenko23 May 2026

Problem

A classifieds marketplace has to recommend listings from about 180 million items to about 6 million users. The signals are click and contact histories, and a good listing can be bought and closed within three days. A two-stage recommender (candidate generation, then ranking) has two typical failure modes. Adding candidate sources or positives the generators never find can stop the metric from growing. A bigger candidate pool can even make the ranker worse.

Example from the article: the Avito ML Cup 2026 recommendation task. The winning team raised the leaderboard score from about 10 (co-visitation alone) to 17.65 with three candidate generators, a CatBoost ranker and transformer features.

Solution in brief

Generate candidates with cheap graph methods (co-visitation, BM25-normalized co-visitation, RP3beta). Rank them with CatBoost trained only on the pool those generators produce. Feed it two kinds of features: which generator found the item and with what score, and ranks of item attributes predicted by a small multitask transformer.

Questions this case answers
  • How to build a two-stage recommender for a very large catalog where the reranker only sees candidates from its own generators and adding candidate sources stops improving the metric
  • How to add sequence-model signals to a gradient-boosting ranker as coarse learnable targets instead of exact next-item predictions
  • How to build a two-stage recommender with multi-source candidate generation and a boosting reranker
  • How to choose the candidate pool size for a gradient-boosting ranker

Tools and methods

  • Co-visitation, BM25 normalization, RP3beta — three candidate generators over user and item interaction graphs
  • Closed-category and item-dynamics heuristics — rule-based re-ranking baseline without a model
  • CatBoost ranker (pairwise) — reranks the merged pool of candidates
  • Multitask transformer — predicts category, region and similar attributes of the next contact
  • Rank-based features — pass transformer output to boosting as ranks instead of logits
Continue reading — it's free

Google or a one-time email code. No password.