The cold-start problem appears when a system needs data that will only exist after launch. It creates a loop: effective predictions require interaction history, but interaction history requires a functioning product. section 6.5
The practical response is approximation rather than waiting for perfect data. Choose sources or interim behavior close enough to the target problem, then replace or refine them as real usage accumulates. section 6.5
Useful ways to break the loop include:
- Generate synthetic examples with computer graphics or other simulation techniques. section 6.5
- Collect, scrape, or purchase relevant public data, accepting that it may not perfectly match the deployment environment. section 6.5
- Create hybrid data by modifying deployment-like examples, or organize a targeted photo session. section 6.5
- Use a naive baseline or heuristic to provide initial value while the system gathers evidence. section 6.5
New products in demand forecasting
New marketplace products lacked historical sales, so the model forecast them poorly. A warming-up heuristic based on sales from products in the same category supplied an initial foundation instead of relying on the model alone. section 9.3.1
The term also describes a serving problem: a serverless inference worker may need several seconds to initialize a large model after sitting idle. Slimmer artifacts, scaling rules, or provider features can reduce that delay, but warm and cold latency may differ sharply. section 15.3.2
