Winner's Solution Walkthrough: 2nd Place (AvitoTech ML Cup 2026)

Alexey Kolegov23 May 2026

Problem

A team building recommendations from marketplace event logs has a very sparse user-item matrix. In such data the links between items matter most, and standard retrievers (ALS, iALS, SVD, ANN, a basic two-tower model, SLIM) did not give a usable result. Only CoVisit, which finds items that often occur together, reached 0.08 on validation. CoVisit looks only at direct item pairs and describes too little of the graph statistics, so a stronger retriever was needed.

Example from the article: in the AvitoTech ML Cup 2026 (the speaker's team placed second), a user who bought a camera can be offered a camera stand because the two items co-occur.

Solution in brief

Replace CoVisit with Pixie, a graph retriever from a 2018 Pinterest paper. It walks a user-item graph (the user, their items, other users, their items), keeps the most relevant users and items, and adds them to the candidate pool. Tune its parameters with popularity corrections, event weights, time decay and pool-size settings, then rerank with a two-stage LightGBM model.

Questions this case answers
  • How to build high-recall candidate generation from very sparse user-item interactions when direct item co-occurrence is too weak
  • How to tune retrieval parameters quickly with a quadratic regression over past runs
  • How to build high-recall candidate generation for sparse user-item interaction data
  • How to use a user-item graph (Pixie) for retrieval when co-visitation is too weak

Tools and methods

  • Pixie — graph candidate generation over a user-item graph
  • CoVisit — baseline retriever, 0.08 on validation
  • P3-alpha, RP3-beta — popularity corrections
  • Quadratic polynomial regression — finds the optimum of a parameter from runs already made
  • Two-stage LGBM reranker — ranks the candidate pool
Continue reading — it's free

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