Design Cold Start for a Large, Turning-Over Creative Catalogue

A messaging platform wants to personalize notification creative selection (not just send-time) across a catalogue of roughly 20,000 creative templates that marketing teams add and retire continuously, at a rate of about 200 new templates per week.

1. Explain why a disjoint per-arm LinUCB model (a separate θ per creative) is a poor fit here, quantifying the practical problem.
2. Propose an action-features design: what features would you put on each creative, and how does this change what happens on a new creative's very first impression?
3. Even with action features, argue why some exploration budget is still necessary for new creatives, and propose a concrete mechanism to guarantee it.