Problem
The AvitoTech ML Cup 2026 asked participants to improve a recommendation model from three months of Avito clickstream. Avito has several verticals (goods, auto, real estate), and goods dominate the raw data because they sell faster. A naive split would produce a competition that mostly measures goods, and a target of "the next item" would reward trivial same-session clicks. An open leaderboard also invites rule-breaking: one team can run several accounts.
Example from the article: from the week after the time cutoff the organizers sampled 200 thousand users with a quota of 10 thousand per vertical.
Solution in brief
Cut the data by time, then sample evaluation users with per-vertical quotas, but build the target from the user's real future. The target is contacts made after a 12-hour gap, minus items the user already clicked. The same team then compared submissions to catch accounts that shared solutions.
Questions this case answers
- How to build a realistic offline evaluation set from clickstream when one segment dominates the raw data
- How to define a recommendation target that rewards long-term interest rather than the next click
- How to detect copied model submissions in a competition from their error patterns
- How to build a realistic offline evaluation set for a recommender from clickstream when verticals are heavily imbalanced
Tools and methods
- Time cutoff — keeps future events out of training and avoids leakage.
- Per-vertical user quotas — rebalance evaluation users across verticals without trimming the target.
- 12-hour gap and contact-based target — long-term interest instead of in-session prediction.
- Recall@160 on candidate generation — the competition metric.
- Submission correlation analysis — finds teams that copy each other.
Google or a one-time email code. No password.