The “cold start” problem — how do you recommend anything to a brand-new user, or recommend a brand-new item with zero interaction history yet — is one of recommendation systems’ most persistent, genuinely fundamental challenges.
The New User Problem
Collaborative filtering (recommending based on what genuinely similar users liked) has literally nothing to work with for a user who just signed up moments ago — no rating or interaction history exists yet at all. Common mitigations include an onboarding preference survey, showing genuinely popular items broadly as a reasonable default starting point, or leveraging any available demographic/contextual signals as a rough initial proxy for taste.
The New Item Problem
A brand-new product with zero purchase/rating history can't be recommended via pure collaborative filtering (no one has interacted with it yet at all to establish a genuine similarity signal). Mitigation: use CONTENT-BASED features instead for new items - category, description, price range, genuinely similar existing items - until enough real interaction data accumulates to switch over to collaborative signals for that specific item.
Modern Approaches
Many production systems now use hybrid models that genuinely blend content-based features (which work fine immediately for new items with zero history) with collaborative signals (which genuinely improve as real interaction data accumulates over time) — smoothly transitioning an item or user from a cold-start, low-confidence state toward increasingly personalized recommendations as more real, genuine interaction data becomes available.
Some systems also deliberately use exploration strategies (occasionally showing a slightly random or genuinely under-explored item, even at a small potential short-term cost to expected relevance) specifically to more quickly gather the interaction data needed to solve their own cold-start problem for new items faster.