합성 데이터

Start with the purpose. Interface test records need valid shapes and edge cases; a training dataset may need meaningful relationships and rare conditions. A dataset suitable for checking a form is not necessarily suitable for estimating population statistics. Document how the data was produced and what real information influenced it. Rule-based generation, simulation, statistical sampling, and generative models create different kinds of errors. Keep generated records distinguishable from observed records in data lineage. Evaluate utility for the specific downstream task. Compare results on an independent real-world test set where appropriate, and inspect subgroup coverage. Synthetic records can amplify a generator’s assumptions or omit uncommon situations even when the overall distribution looks plausible. Evaluate privacy separately. A generator may reproduce information from its source data, and removing obvious identifiers is not a universal privacy guarantee. Differential privacy is one formal framework, but its guarantees depend on the actual mechanism and parameters. Review claims about privacy and utility independently rather than assuming one implies the other.