1. Cold start and exploration

"What does your model do for a user who signed up five minutes ago?" Every personalization model learns from history: who opened what, who clicked what. Cold start is the moment it has to act on a user, an item, or a whole market with no history at all. The answer has two halves:

• First, a fallback that makes a sensible choice right now from whatever you still know.
• Second, exploration: deliberately spending a small share of traffic on choices the model is unsure about, so that the missing history gets collected.

We will use one running example. A food-delivery app sends each user at most one push notification a day, and a model picks which of about 12 notification types to send: a reorder reminder, a promo, a new-restaurant alert, and so on. Success is the open rate, typically around 4%. Exploring here is cheap, because a wasted send costs almost nothing.

That is not true everywhere: a payments processor scoring fraud for a merchant onboarded today cannot approve risky transactions at random to learn, so it stops at the fallback and lets ordinary traffic build history. Back in the food-delivery app, new users sign up every day, the product team keeps adding notification types, and the app keeps launching in new cities. The notification model therefore encounters all three kinds of cold start.

In a machine learning (ML) system design round this is usually a 3-5 minute follow-up after the main model is designed, not a whole round. Mid-level candidates are expected to name what is new and give a fallback with a prior. Senior candidates are expected to explain why a greedy model never learns about new items, pick an exploration policy, cap its cost, and log the data so the next model can use it. By the end you should be able to answer "what happens on day one?" for any new prompt.