Winner's Solution Walkthrough: 3rd Place (AvitoTech ML Cup 2026)

Sergey Arefyev23 May 2026

Problem

A team has a log of user interactions with millions of items and must predict which items each user will interact with next. A full item-to-item matrix does not fit: about 10 million items, about 4 million after filtering. Plain co-interaction counts are also dominated by popular items and ignore when the interaction happened.

Example from the article: AvitoTech ML Cup 2026, an Avito recommendation competition across 7 verticals. The speaker joined about a week before the end, reused a pipeline from earlier competitions and took 3rd place.

Solution in brief

Build a candidate generator from item-to-item co-interaction matrices, one matrix per category rather than one for the whole catalog. Weight it by item popularity and interaction recency, and combine the category matrices per user. Rerank the candidates with CatBoost on matrix-derived and user-level time features.

Questions this case answers
  • How to generate next-interaction candidates from item-to-item co-occurrence when the catalog is too large for a full matrix
  • How to down-weight popular items and old interactions in co-occurrence counts for next-item prediction
  • How to generate candidates for next-purchase prediction from item co-occurrence when the catalog has millions of items
  • How to keep an item-to-item matrix tractable on a catalog of about 4 million items

Tools and methods

  • Item-to-item co-interaction matrix — cell value is the number of users who interacted with both items (interacted, not only bought).
  • Per-category matrices — 50 matrices instead of one, the largest quality gain in the talk.
  • Time-decay weight — a power of (1+age)/14 applied inside the matrix.
  • CatBoost reranker — scores candidates using 60 and 90 features, 5 validation folds.
  • Claude Code — ran the grid search, ported the computation to GPU and added extra features.
Continue reading — it's free

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