Cold start is the recommendation problem that appears when a system has too little history to make confident predictions. The most common versions are a new user with almost no interaction data, a new item with little or no engagement history, or a new context where past patterns do not transfer cleanly. In recommendation systems, cold start matters because many strong ranking methods depend on previous behavior that may not exist yet.
Why It Happens
A recommender often learns from clicks, listens, purchases, ratings, follows, or other feedback. When that history is missing, the system has to rely more on metadata, editorial choices, onboarding questions, popularity, content similarity, or other fallback strategies. That is why cold start is closely related to candidate generation, embeddings, and model evaluation: the system still needs a reasonable way to retrieve and score items before it has strong behavioral evidence.
Why It Matters
Cold start matters because it can shape what new users or new creators ever get a chance to see. If the system handles cold start poorly, new users receive weak recommendations and may leave, while new items struggle to reach the right audience. If the system handles it well, discovery improves and the catalog feels much more alive.
Where You See It
Cold start appears in music services trying to recommend a new artist, in e-commerce when a new shopper or product enters the catalog, and in social feeds when a new creator posts without much historical engagement. Strong systems often mix popularity, metadata, editorial signals, and exploration to avoid waiting too long for feedback to accumulate.
Related Yenra articles: Music Recommendation Services and E-Commerce Recommendation Engines.
Related concepts: Recommender System, Candidate Generation, Embedding, Model Evaluation, and Model Monitoring.