How We Relaunched Recommendations at Auto.ru

Vadim Kokhtev27 May 2023

Problem

Avto.ru, a large Russian car classifieds service, had only textbook item-based collaborative filtering: similar listings on a car page and email selections. New logic was hard to add. The team wanted an infinite personalized feed, measured by time in the app rather than calls or chat messages. Two things made that hard: a listing is a real car that vanishes once sold, and a model trained on its own output sees only what it already chose to show.

Example from the article: a pure ALS model on listing vectors learned popularity, so a cheap Chinese crossover came out as "similar" to a Maybach worth tens of millions of rubles.

Solution in brief

Test the idea with a feed built on the existing parametric search, then build candidate generation and ranking. Candidates come from business rules first, then embeddings. Training data comes from random ranking shown to a share of users. The ranker learns time on service after impression; the same ranker over catalog-style candidates replaced "similar listings".

Questions this case answers
  • How to build a personalized feed of unique items that expire once they are taken, when collaborative similarity signals are too thin to rely on
  • How to collect unbiased training data for a ranker that only ever sees the items it has already chosen to show
  • How to build a personalized infinite feed for a classifieds service where listings disappear quickly
  • How to collect unbiased training data when a recommender only shows what it chose itself

Tools and methods

  • Candidate rules — popular, expensive, old, previously viewed, same brand
  • Item embeddings (ALS) in a KNN index — approximate-search candidate source
  • Random ranking on a share of users — breaks the feedback loop and collects training data
  • Pointwise ranking model — target is time on service after impression; pairwise losses suit clicks
  • Proportional mixing — every Nth feed item is a review or article
Continue reading — it's free

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