Why feeds get stuck on popular items

A recommendation model learns from clicks, and users can only click what the feed shows them. That closes a loop: items the model already likes get shown, collect clicks and stay; items it has never shown collect nothing, so the model never finds out whether they are good.

1. 1
   The chart is a seeded simulation of 300 users and 1,000 items. Each round, every user sees 10 items and clicks each one with its true click probability. The model starts from a two-tower prediction for every user and item pair, and updates its estimate for a pair only when that item is shown to that user. Items it never shows keep their starting estimate forever.
2. 2
   With no exploration, the feed shows each user the 10 items with the highest estimates. After 40 rounds only 8% of the catalog has ever been shown. The impression Gini, which is 0 when every item is shown equally and 1 when one item gets every impression, is 0.97. Click rate still climbs to 46.2%, so the dashboard looks healthy while the logs cover a small, mostly popular corner of the catalog.
3. 3
   With 10% exploration, each slot has a 1 in 10 chance of going to a random item instead. Every item is shown by round 22 and the Gini falls to 0.89. The price is visible too: in round 40 the click rate is 42.1%, 4.1 points lower, because random items are usually less relevant than the model's picks.