Chasing All the Hares: Managing Different Content Types in One Recommender System

Anastasia Zinina24 May 2025

Problem

A bank's retail mobile app shows up to about 50 offers in banners and sliders, and the offers serve different goals: product sales, engagement (stories, quizzes), retention promos, transfers. Content was loaded by business rules, so every change meant meetings, because each team wanted its offer on top. The team needed one automated ranking for all content types, including new offers with no click history.

Example from the article: a new "super cashback from a partner" offer attracts clicks but barely affects money or customer lifetime value (CLV), yet it takes impressions from permanent offers that do.

Solution in brief

Rank by predicted click as a common proxy. A BERT encoder over the customer's event sequence feeds a classifier that scores permanent ("hot") offers. New ("cold") offers get Thompson Sampling with per-cluster beta distributions, and hot scores become narrow beta distributions so both groups are ranked together. A quantile rule guarantees a share of top positions to important hot offers, which blocks clickbait.

Questions this case answers
  • How to rank items of different types with different business metrics in one model, including new items with no history
  • How to stop low-value items from taking exposure away from items that drive the business metrics
  • How to rank banner offers of different content types by one model when each type has its own business metric
  • How to rank new (cold-start) offers together with proven ones when there is no history to train on

Tools and methods

  • Lightweight BERT event encoder — 17M parameters, masked-event pretraining, [CLS] vector as the customer embedding
  • Two-layer classifier — probabilities over about 20 banners
  • Thompson Sampling, beta distributions — cold offers, statistics per customer cluster
  • Distribution of the maximum, N% quantile — the "top guarantee" for hot offers
Continue reading — it's free

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