Evaluation Framework

SDV contains a Synthetic Data Evaluation Framework that facilitates the task of evaluating the quality of your Synthetic Dataset by applying multiple Synthetic Data Metrics on it and reporting results in a comprehensive way.

Using the SDV Evaluation Framework

To evaluate the quality of synthetic data we essentially need two things: real data and synthetic data that pretends to resemble it.

Let us start by loading a demo table and generate a synthetic replica of it using the GaussianCopula model.

In [1]: from sdv.demo import load_tabular_demo

In [2]: from sdv.tabular import GaussianCopula

In [3]: real_data = load_tabular_demo('student_placements')

In [4]: model = GaussianCopula()

In [5]: model.fit(real_data)

In [6]: synthetic_data = model.sample()

After the previous steps we will have two tables:

  • real_data: A table containing data about student placements

In [7]: real_data.head()
Out[7]: 
   student_id gender  second_perc  high_perc high_spec  degree_perc degree_type  work_experience  experience_years  employability_perc mba_spec  mba_perc   salary  placed start_date   end_date duration
0       17264      M        67.00      91.00  Commerce        58.00    Sci&Tech            False                 0                55.0   Mkt&HR     58.80  27000.0    True 2020-07-23 2020-10-12      3.0
1       17265      M        79.33      78.33   Science        77.48    Sci&Tech             True                 1                86.5  Mkt&Fin     66.28  20000.0    True 2020-01-11 2020-04-09      3.0
2       17266      M        65.00      68.00      Arts        64.00   Comm&Mgmt            False                 0                75.0  Mkt&Fin     57.80  25000.0    True 2020-01-26 2020-07-13      6.0
3       17267      M        56.00      52.00   Science        52.00    Sci&Tech            False                 0                66.0   Mkt&HR     59.43      NaN   False        NaT        NaT      NaN
4       17268      M        85.80      73.60  Commerce        73.30   Comm&Mgmt            False                 0                96.8  Mkt&Fin     55.50  42500.0    True 2020-07-04 2020-09-27      3.0
  • synthetic_data: A synthetically generated table that contains data in the same format and with similar statistical properties as the real_data.

In [8]: synthetic_data.head()
Out[8]: 
   student_id gender  second_perc  high_perc high_spec  degree_perc degree_type  work_experience  experience_years  employability_perc mba_spec   mba_perc        salary  placed start_date   end_date duration
0       17313      M    41.068012  56.085381  Commerce    44.883622   Comm&Mgmt            False                 1           50.014502   Mkt&HR  50.329875  24956.891559    True 2020-04-05 2021-04-11     12.0
1       17459      M    83.551016  65.519349  Commerce    67.995809    Sci&Tech            False                 1           77.288552  Mkt&Fin  66.383096  31318.783931    True 2020-01-25 2020-07-03      NaN
2       17459      F    71.216682  68.833595  Commerce    85.226367   Comm&Mgmt            False                 0           54.276205  Mkt&Fin  67.559014  25514.777281    True 2020-01-08 2020-02-05      3.0
3       17298      M    80.828939  76.962830  Commerce    73.777104   Comm&Mgmt            False                 1           75.191165  Mkt&Fin  78.558689  35690.903886    True 2020-01-11 2020-06-09      3.0
4       17440      F    63.275924  62.635678  Commerce    61.609046   Comm&Mgmt            False                 1           65.422342  Mkt&Fin  71.237561  32701.627256    True 2020-01-06 2020-01-06      3.0

Note

For more details about this process, please visit the GaussianCopula Model guide.

Computing an overall score

The simplest way to see how similar the two tables are is to import the sdv.evaluation.evaluate function and run it passing both the synthetic_data and the real_data tables.

In [9]: from sdv.evaluation import evaluate

In [10]: evaluate(synthetic_data, real_data)
Out[10]: 0.6093353724148053

The output of this function call will be a number between 0 and 1 that will indicate how similar the two tables are, being 0 the worst and 1 the best possible score.

How was the obtained score computed?

The evaluate function applies a collection of pre-configured metric functions and returns the average of the scores that the data obtained on each one of them. In most scenarios this can be enough to get an idea about the similarity of the two tables, but you might want to explore the metrics in more detail.

In order to see the different metrics that were applied you can pass and additional argument aggregate=False, which will make the evaluate function return a dictionary with the scores that each one of the metrics functions returned:

In [11]: evaluate(synthetic_data, real_data, aggregate=False)
Out[11]: 
                    metric                                     name  raw_score  normalized_score  min_value  max_value      goal
1        LogisticDetection             LogisticRegression Detection   0.339336      3.393359e-01        0.0        1.0  MAXIMIZE
2             SVCDetection                            SVC Detection   0.406922      4.069220e-01        0.0        1.0  MAXIMIZE
11         GMLogLikelihood           GaussianMixture Log Likelihood -42.161881      4.890215e-19       -inf        inf  MAXIMIZE
12                  CSTest                              Chi-Squared   0.896715      8.967145e-01        0.0        1.0  MAXIMIZE
13                  KSTest  Inverted Kolmogorov-Smirnov D statistic   0.928488      9.284884e-01        0.0        1.0  MAXIMIZE
14          KSTestExtended  Inverted Kolmogorov-Smirnov D statistic   0.921395      9.213953e-01        0.0        1.0  MAXIMIZE
27  ContinuousKLDivergence   Continuous Kullback–Leibler Divergence   0.581344      5.813441e-01        0.0        1.0  MAXIMIZE
28    DiscreteKLDivergence     Discrete Kullback–Leibler Divergence   0.828046      8.280459e-01        0.0        1.0  MAXIMIZE

Can I control which metrics are applied?

By default, the evaluate function will apply all the metrics that are included within the SDV Evaluation framework. However, the list of metrics that are applied can be controlled by passing a list with the names of the metrics that you want to apply.

For example, if you were interested on obtaining only the CSTest and KSTest metrics you can call the evaluate function as follows:

In [12]: evaluate(synthetic_data, real_data, metrics=['CSTest', 'KSTest'])
Out[12]: 0.9126014582897475

Or, if we want to see the scores separately:

In [13]: evaluate(synthetic_data, real_data, metrics=['CSTest', 'KSTest'], aggregate=False)
Out[13]: 
   metric                                     name  raw_score  normalized_score  min_value  max_value      goal
0  CSTest                              Chi-Squared   0.896715          0.896715        0.0        1.0  MAXIMIZE
1  KSTest  Inverted Kolmogorov-Smirnov D statistic   0.928488          0.928488        0.0        1.0  MAXIMIZE

For more details about all the metrics that exist for the different data modalities please check the corresponding guides.