Cold Start

A new user has no history and a new item has no audience, so the method that works by finding similar behaviour has nothing to work from. Every catalogue has a long tail of items in this state permanently, and a system that only recommends what it has evidence for will never gather evidence about anything else.

The fix is to fall back on content — text, images, categories — which is why hybrid systems exist rather than being an improvement on pure collaborative filtering. The subtler problem is that cold start is self-reinforcing: an item never shown gathers no interactions, so exploration has to be deliberate and paid for, at a measurable cost in short-term engagement.

With no interactions, a collaborative model's item factor is at its prior, so the score is uninformative rather than low — the model has no opinion and the ranking treats that as an opinion. Content features supply an initial estimate; the alternative is an exploration bonus that decays with interaction count, which buys evidence at a known price.