Metrics
We assess the quality of synthetic data both qualitatively through visualizations and quantitatively using metric results.
Evaluate with Metrics
For metric functions, please refer to rockfish.labs.metrics.
Distribution Metrics
| Metric | Metric Value Definition |
|---|---|
| Range coverage | 1 = Best; 0 = Worst |
| Category coverage | 1 = Best; 0 = Worst |
| Total Variation distance | 0 = Best; 1 = Worst |
| Jensen-Shannon distance | 0 = Best ; 1 = Worst |
| Wasserstein distance | 0 = Best; value at positive infinity = Worst |
| Kolmogorov-Smirnov distance | 0 = Best; 1 = Worst |
Correlation Metric
Pearson correlation
- Correlation coefficient between -1 and 1
- -1 means perfectly negative linear relationship between two numerical fields
- 0 means no relationship between two numerical fields
- 1 means perfectly positive relationship between two numerical fields
- P value: if p value is low (generally <0.05), the correlation coefficient is statistically significant.
Association Metric
Cramér's V
When correction = True, Cramér's V includes bias correction (please refer to this research to understand the bias correction). When correction=False, bias correction is not applied. By default, bias correction is enabled.
- 0 means no association between two categorical fields
- 1 means perfect association between two categorical fields