Make your synthesis reproducible

In the previous example, we created our first synthetic dataset. You may have noticed that we created the Synthesiser using the random_seed parameter:

synthesiser = Synthesiser(random_seed=1)

Without specifying a random_seed, if you run the same code again, you may notice that the generated synthetic dataset is different each time. This is expected. Synthetic data generation is a stochastic process: many synthesis methods involve randomness in, for example, sampling.

Sometimes this is exactly what you want. However, when developing analyses, writing reports or collaborating with others, it is often useful to generate exactly the same synthetic dataset every time. This is called reproducibility.

Generate two synthetic datasets

Let’s start by fitting a Synthesiser without specifying a random seed.

from sklearn.datasets import load_diabetes
from synthpop import Synthesiser

data = load_diabetes(as_frame=True).frame

synthesiser = Synthesiser()
synthesiser.fit(data)

synthetic_data_1 = synthesiser.generate()
synthetic_data_2 = synthesiser.generate()

Both datasets were generated from the same fitted Synthesiser, but they are not identical.

Note: the results below may vary from yours as they are not reproducible.

synthetic_data_1.head(3)

	age	sex	bmi	bp	s1	s2	s3	s4	s5	s6	target
0	-0.107226	-0.0446416	-0.011595	-0.0400989	-0.0153285	-0.00131388	-0.0323559	0.0158583	-0.027129	0.0196328	55
1	0.0489735	-0.0446416	-0.0256066	-0.00567042	0.0176944	0.0434664	0.0707299	-0.0247329	-0.0159989	0.0610539	113
2	-0.0563701	-0.0446416	0.0444512	-0.0194418	-0.00707277	-0.0238606	-0.0139477	-0.00259226	0.0757406	-0.013504	121

synthetic_data_2.head(3)

	age	sex	bmi	bp	s1	s2	s3	s4	s5	s6	target
0	0.0671362	0.0506801	-0.0148285	0.0586076	0.0176944	0.0494162	-0.00658447	-0.00259226	-0.0332456	0.0569118	111
1	-0.0273098	0.0506801	-0.00728377	-0.0297704	0.0465894	0.0732155	-0.0286743	0.0343089	-0.00149595	-0.0466409	200
2	0.0417084	0.0506801	-0.0223731	-0.0400989	-0.030464	0.00119131	-0.021311	0.0158583	-0.0109033	0.0444855	83

Although both datasets preserve similar statistical properties, the individual synthetic records differ because new random samples are drawn each time.

synthetic_data_1.describe()

	age	sex	bmi	bp	s1	s2	s3	s4	s5	s6	target
count	442	442	442	442	442	442	442	442	442	442	442
mean	0.000345173	0.00150962	0.000777878	-0.00262506	-0.00255579	-0.00409148	-0.00128691	-0.00179496	0.0023855	-0.0012089	151.928
std	0.0484934	0.0476909	0.0485026	0.0476364	0.0488929	0.049221	0.0486984	0.0477245	0.0468386	0.0464237	76.3776
min	-0.107226	-0.0446416	-0.0891975	-0.112399	-0.126781	-0.115613	-0.102307	-0.0763945	-0.104366	-0.129483	31
25%	-0.0409318	-0.0446416	-0.0331513	-0.0366561	-0.0373437	-0.0356819	-0.0360376	-0.0394934	-0.0307479	-0.033179	89
50%	0.00538306	-0.0446416	-0.00620595	-0.0108347	-0.00707277	-0.010082	-0.00658447	-0.00259226	0.00200444	0.00306441	137.5
75%	0.0380759	0.0506801	0.0315175	0.0287581	0.0245741	0.0262432	0.0302319	0.0343089	0.0354587	0.0279171	202
max	0.110727	0.0506801	0.170555	0.125158	0.153914	0.198788	0.173816	0.185234	0.133597	0.135612	341

synthetic_data_2.describe()

	age	sex	bmi	bp	s1	s2	s3	s4	s5	s6	target
count	442	442	442	442	442	442	442	442	442	442	442
mean	0.000575288	0.00021566	0.00363823	0.001088	1.55649e-05	-0.00141909	-0.00410228	0.00398565	0.00285974	0.00231471	155.394
std	0.0492173	0.0476323	0.047053	0.0482004	0.0471131	0.0470191	0.0447474	0.0500362	0.0491795	0.0484639	77.315
min	-0.107226	-0.0446416	-0.0902753	-0.108956	-0.108893	-0.112795	-0.102307	-0.0763945	-0.126097	-0.137767	31
25%	-0.0418399	-0.0446416	-0.0320734	-0.0366561	-0.0345918	-0.0338813	-0.0397192	-0.0394934	-0.0316777	-0.0300724	88.25
50%	0.0090156	-0.0446416	-0.0013558	-0.00567042	-0.00363289	-0.00976889	-0.0139477	-0.00259226	0.00286131	0.000993356	151
75%	0.0380759	0.0506801	0.0358287	0.0390866	0.0256061	0.0293747	0.0265503	0.0343089	0.037667	0.0279171	216.75
max	0.0961965	0.0506801	0.137143	0.125158	0.153914	0.155887	0.159089	0.185234	0.133597	0.135612	341

Set a random seed

To obtain reproducible results, specify the random_seed parameter when creating the Synthesiser.

synthesiser_1 = Synthesiser(random_seed=1)

synthesiser_1.fit(data)

synthetic_data_1 = synthesiser_1.generate()

Now create a new Synthesiser with the same random seed.

synthesiser_2 = Synthesiser(random_seed=1)

synthesiser_2.fit(data)

synthetic_data_2 = synthesiser_2.generate()

Both datasets are now identical.

synthetic_data_1.equals(synthetic_data_2)

True

Because every random decision is made from the same initial random state, the complete synthesis process can be reproduced exactly.

When should you use a random seed?

Using a random seed is recommended when you want reproducible results, for example when:

• writing documentation or tutorials;

• comparing different synthesis methods;

• debugging your synthesis pipeline;

• sharing analyses with colleagues;

• running automated tests.

Reproducibility and multiple synthetic datasets

Sometimes you may want several different synthetic datasets from the same fitted Synthesiser, for example to quantify the variability introduced by the synthesis process.

However, if a random seed is provided, you cannot simply call generate() multiple times as we saw in the first code block of this page.

synthesiser = Synthesiser(random_seed=1)
synthesiser.fit(data)

synthetic_data_1 = synthesiser.generate()
synthetic_data_2 = synthesiser.generate()

synthetic_data_1.equals(synthetic_data_2)

True

To create multiple reproducible synthetic dataset from the same fitted synthesiser, you can use the random_seed parameter in the generate() function.

synthetic_data_3 = synthesiser.generate(random_seed=2)

Each generated dataset with a different random seed will be different, because the synthesiser continues drawing new random samples.

synthetic_data_3.equals(synthetic_data_1)

False

Omitting the random_seed parameter will again provide you with the first synthetic dataset.

synthetic_data_4 = synthesiser.generate()

synthetic_data_4.equals(synthetic_data_1)

True

Next steps

Now that your synthesis is reproducible, the next examples in this module show you how to generate larger datasets and how to choose a synthesis order.