A Neural Network User Interest Model: How We Improved Diversity and Relevance of Recommendations on the Avito Home Page

Yaroslav Khnykov24 May 2025

Problem

Avito's home page is an endless personalized feed of listings. It carries about 50% of all listing views and about 30% of all buyer contacts with sellers. The system is two-stage: five candidate generators, then a boosting ranker. The ranker scores each item independently, so similar items get similar scores and clump together after sorting. With thousands of candidates, the whole top can fall into one category. Users see a monotonous feed, miss their other interests and may leave.

Example from the article: a user browsing iPhones gets a block of near-identical iPhone listings, although they may not need a phone at all right now.

Solution in brief

Add a diversification layer, the "blender", on top of the ranker. It samples categories in proportion to the user's interest in each and takes the best remaining listing from the chosen category. Interest first came from a decayed event counter, then from a small Transformer Encoder over the user's history that predicts category proportions.

Questions this case answers
  • How to diversify a ranked feed of items so that near-duplicate items do not clump together when the ranker scores each item independently
  • How to estimate a user's interest in each category from their event history and sample categories in proportion to that interest so the feed stays both relevant and diverse
  • How to increase the diversity of a ranked recommendation feed without losing relevance
  • How to estimate a user's interest in each product category from their event history

Tools and methods

  • Blender — mixes ranked listings by category, sampling categories by user interest
  • Decayed event counter — baseline interest per category: event weights times a time decay
  • Transformer Encoder with Attention Pooling — predicts the share of future favorites and contacts by category
  • Temperature softmax — smooths model outputs to trade relevance for diversity
Continue reading — it's free

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