Cold-start policy

Exploration rate (epsilon): 15%

This case study models an epsilon-greedy recommendation engine choosing among 1,500 catalog items, some brand new with no click history. Higher exploration shows more new/unproven items to build data on them, at the cost of some clicks lost to items less certain to perform well right now.

Week outcome

Overall click-through rate—

New items that got any exposure—

Catalog items never shown—

Impression share by item rank (purple = established items, teal = new items)