Improving Performance Summary Metric Scores
If your privacy or quality metrics are lower than expected or required, we have the following recommendations:
| Potential cause | Explanation | Solution |
|---|---|---|
| Input data quality | Synthetic data models learn patterns directly from the input data. If the original dataset contains errors, inconsistencies, extreme noise, or poorly defined values, these issues may be reflected or amplified in the synthetic output, leading to lower quality metrics. | Clean and validate the input data by correcting obvious errors, standardising formats, and removing invalid or out-of-scope values before generation. |
| Dataset size | Synthetic data generation requires sufficient input data to train the model. An adequate number of rows relative to columns is needed for the model to reliably capture statistical attributes. | Either increase the number of rows or reduce the number of columns. Generally, we recommend for the ratio of Rows: Columns to be minimally 50:1 and ideally > 200:1. Find out how here. |
| Feature relevance and column selection | Even if you have a large enough dataset size (i.e., many rows), including columns that are irrelevant or weakly related to the analysis objective can dilute the model’s ability to learn meaningful patterns, reducing quality. | When preparing your real data or in Step 2, focus on columns that directly support your use case and exclude the rest. |
| Model selection | The selected model may not be well suited to your data or use case. For example:
|
Try another model that aligns to your data characteristics, analysis goals, and privacy requirements. Find out how in the Model Selection section. |
| Model hyperparameter selection | Even with the right model, unsuitable hyperparameters can prevent it from learning your data effectively. Hyperparameters control how the model trains (e.g., how long it trains, how much data it processes at once, and how quickly it updates). For example, in DDPM, too few training iterations may result in poor distributional fit, while overly aggressive learning rates can lead to unstable training. |
Iteratively adjust hyperparameters. Find out how on the Custom page. |
Need help? Try our Support Assistant or visit the Support page.
