Not the most accurate, the most appropriate
Production model choice balances accuracy against latency, interpretability, cost, and operational burden.
A practical ladder
Climb only as far as the problem demands.
- Heuristics rules you can ship in a day as a baseline
- Linear and tree models fast, interpretable, strong on tabular data
- Gradient boosted trees often the best tabular accuracy per dollar
- Deep networks needed for text, images, audio, and huge data
Decision factors
- Latency a deep model may blow the inference budget
- Interpretability regulated domains may require explainable models
- Training cost retraining a large model frequently is expensive
- Data volume deep nets need lots of data to beat boosting
Key idea
Pick the simplest model that clears your constraints. Complexity is a cost you pay forever in latency, dollars, and debugging.