How does cold start problem in ML systems relate to bias and variance?

Cold start increases bias due to lack of data and increases variance due to unreliable early predictions.

The cold start problem occurs when a model has insufficient data for new users, items, or contexts. In this scenario, models rely on global averages or heuristic rules, leading to high bias because personalization is weak.

At the same time, sparse signals cause unstable predictions across similar inputs, increasing variance. Systems often combine content-based features, metadata priors, and transfer learning to mitigate this.

Architecturally, cold start is handled using fallback models, hybrid recommenders, and embedding initialization from pre-trained models.