CTGAN Model

In this guide we will go through a series of steps that will let you discover functionalities of the CTGAN model, including how to:

  • Create an instance of CTGAN.

  • Fit the instance to your data.

  • Generate synthetic versions of your data.

  • Use CTGAN to anonymize PII information.

  • Customize the data transformations to improve the learning process.

  • Specify hyperparameters to improve the output quality.

What is CTGAN?

The sdv.tabular.CTGAN model is based on the GAN-based Deep Learning data synthesizer which was presented at the NeurIPS 2020 conference by the paper titled Modeling Tabular data using Conditional GAN.

Let’s now discover how to learn a dataset and later on generate synthetic data with the same format and statistical properties by using the CTGAN class from SDV.

Quick Usage

We will start by loading one of our demo datasets, the student_placements, which contains information about MBA students that applied for placements during the year 2020.

Warning

In order to follow this guide you need to have ctgan installed on your system. If you have not done it yet, please install ctgan now by executing the command pip install sdv in a terminal.

In [1]: from sdv.demo import load_tabular_demo

In [2]: data = load_tabular_demo('student_placements')

In [3]: data.head()
Out[3]: 
   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

As you can see, this table contains information about students which includes, among other things:

  • Their id and gender

  • Their grades and specializations

  • Their work experience

  • The salary that they where offered

  • The duration and dates of their placement

You will notice that there is data with the following characteristics:

  • There are float, integer, boolean, categorical and datetime values.

  • There are some variables that have missing data. In particular, all the data related to the placement details is missing in the rows where the student was not placed.

Let us use CTGAN to learn this data and then sample synthetic data about new students to see how well de model captures the characteristics indicated above. In order to do this you will need to:

  • Import the sdv.tabular.CTGAN class and create an instance of it.

  • Call its fit method passing our table.

  • Call its sample method indicating the number of synthetic rows that you want to generate.

In [4]: from sdv.tabular import CTGAN

In [5]: model = CTGAN()

In [6]: model.fit(data)

Note

Notice that the model fitting process took care of transforming the different fields using the appropriate Reversible Data Transforms to ensure that the data has a format that the underlying CTGANSynthesizer class can handle.

Generate synthetic data from the model

Once the modeling has finished you are ready to generate new synthetic data by calling the sample method from your model passing the number of rows that we want to generate.

In [7]: new_data = model.sample(200)

This will return a table identical to the one which the model was fitted on, but filled with new data which resembles the original one.

In [8]: new_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       17424      M    61.699746  83.718700  Commerce    70.280676    Sci&Tech            False                 0           58.433171  Mkt&Fin  70.499948           NaN    True 2020-06-15 2020-07-31      6.0
1       17454      F    28.937288  64.358821   Science    65.522645   Comm&Mgmt            False                 0           95.313676   Mkt&HR  55.380654  60481.489494    True        NaT 2020-08-30      NaN
2       17439      F    58.922229  82.018902      Arts    76.902771    Sci&Tech            False                 0           80.718947  Mkt&Fin  57.571961           NaN    True 2020-01-12        NaT      6.0
3       17414      M    73.880056  74.937830   Science    76.265360   Comm&Mgmt            False                 0           85.179291  Mkt&Fin  48.703399           NaN    True 2020-03-27        NaT     12.0
4       17298      M    69.046830  70.492684   Science    86.207639    Sci&Tech            False                 0           62.793179  Mkt&Fin  59.174083           NaN    True 2020-01-14        NaT      6.0

Note

You can control the number of rows by specifying the number of samples in the model.sample(<num_rows>). To test, try model.sample(10000). Note that the original table only had ~200 rows.

Save and Load the model

In many scenarios it will be convenient to generate synthetic versions of your data directly in systems that do not have access to the original data source. For example, if you may want to generate testing data on the fly inside a testing environment that does not have access to your production database. In these scenarios, fitting the model with real data every time that you need to generate new data is feasible, so you will need to fit a model in your production environment, save the fitted model into a file, send this file to the testing environment and then load it there to be able to sample from it.

Let’s see how this process works.

Save and share the model

Once you have fitted the model, all you need to do is call its save method passing the name of the file in which you want to save the model. Note that the extension of the filename is not relevant, but we will be using the .pkl extension to highlight that the serialization protocol used is pickle.

In [9]: model.save('my_model.pkl')

This will have created a file called my_model.pkl in the same directory in which you are running SDV.

Important

If you inspect the generated file you will notice that its size is much smaller than the size of the data that you used to generate it. This is because the serialized model contains no information about the original data, other than the parameters it needs to generate synthetic versions of it. This means that you can safely share this my_model.pkl file without the risc of disclosing any of your real data!

Load the model and generate new data

The file you just generated can be send over to the system where the synthetic data will be generated. Once it is there, you can load it using the CTGAN.load method, and then you are ready to sample new data from the loaded instance:

In [10]: loaded = CTGAN.load('my_model.pkl')

In [11]: new_data = loaded.sample(200)

Warning

Notice that the system where the model is loaded needs to also have sdv and ctgan installed, otherwise it will not be able to load the model and use it.

Specifying the Primary Key of the table

One of the first things that you may have noticed when looking that demo data is that there is a student_id column which acts as the primary key of the table, and which is supposed to have unique values. Indeed, if we look at the number of times that each value appears, we see that all of them appear at most once:

In [12]: data.student_id.value_counts().max()
Out[12]: 1

However, if we look at the synthetic data that we generated, we observe that there are some values that appear more than once:

In [13]: new_data[new_data.student_id == new_data.student_id.value_counts().index[0]]
Out[13]: 
     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
5         17435      M    73.646986  98.257372      Arts    68.372614   Comm&Mgmt            False                 0          103.240845  Mkt&Fin  52.049758  29332.138210    True 2020-07-11        NaT      3.0
39        17435      M    64.014901  59.092988  Commerce    51.820853   Comm&Mgmt             True                 0           77.100064   Mkt&HR  47.358588  29004.950063    True 2020-03-16 2020-04-08      3.0
94        17435      F    43.963257  62.471194  Commerce    60.002170   Comm&Mgmt             True                 0          102.944376   Mkt&HR  53.425593           NaN    True 2020-04-01 2020-07-13      3.0
129       17435      M    60.780645  68.727590  Commerce    75.071070    Sci&Tech             True                 0           94.857543  Mkt&Fin  45.877561           NaN    True        NaT        NaT     12.0
133       17435      M    72.573262  68.607346  Commerce    81.368647      Others            False                 0           73.839354   Mkt&HR  51.120447           NaN   False 2020-08-13 2020-12-20      3.0

This happens because the model was not notified at any point about the fact that the student_id had to be unique, so when it generates new data it will provoke collisions sooner or later. In order to solve this, we can pass the argument primary_key to our model when we create it, indicating the name of the column that is the index of the table.

In [14]: model = CTGAN(
   ....:     primary_key='student_id'
   ....: )
   ....: 

In [15]: model.fit(data)

In [16]: new_data = model.sample(200)

In [17]: new_data.head()
Out[17]: 
   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           0      F    65.303279  55.046971  Commerce    65.994968   Comm&Mgmt             True                 0          101.689991  Mkt&Fin  83.747836  28776.045186   False        NaT 2020-08-29      6.0
1           1      F    88.865623  84.548175  Commerce    72.082708   Comm&Mgmt             True                 0           63.221083  Mkt&Fin  57.549107           NaN    True 2020-07-24 2020-06-28     12.0
2           2      F    74.502908  60.371752  Commerce    71.002056    Sci&Tech             True                 0           59.251831  Mkt&Fin  69.141486           NaN    True 2020-03-03 2020-07-10      6.0
3           3      M    74.579937  76.262574  Commerce    70.327636   Comm&Mgmt             True                 0           83.870629   Mkt&HR  68.849868           NaN    True        NaT 2020-03-14      NaN
4           4      M    54.085977  74.655591  Commerce    69.084811   Comm&Mgmt             True                 0           56.246279  Mkt&Fin  64.962192  23254.883178   False 2020-05-29 2020-08-19     12.0

As a result, the model will learn that this column must be unique and generate a unique sequence of values for the column:

In [18]: new_data.student_id.value_counts().max()
Out[18]: 1

Anonymizing Personally Identifiable Information (PII)

There will be many cases where the data will contain Personally Identifiable Information which we cannot disclose. In these cases, we will want our Tabular Models to replace the information within these fields with fake, simulated data that looks similar to the real one but does not contain any of the original values.

Let’s load a new dataset that contains a PII field, the student_placements_pii demo, and try to generate synthetic versions of it that do not contain any of the PII fields.

Note

The student_placements_pii dataset is a modified version of the student_placements dataset with one new field, address, which contains PII information about the students. Notice that this additional address field has been simulated and does not correspond to data from the real users.

In [19]: data_pii = load_tabular_demo('student_placements_pii')

In [20]: data_pii.head()
Out[20]: 
   student_id                                            address 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        70304 Baker Turnpike\nEricborough, MS 15086      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    805 Herrera Avenue Apt. 134\nMaryview, NJ 36510      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        3702 Bradley Island\nNorth Victor, FL 12268      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                   Unit 0879 Box 3878\nDPO AP 42663      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  96493 Kelly Canyon Apt. 145\nEast Steven, NC 3...      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

If we use our tabular model on this new data we will see how the synthetic data that it generates discloses the addresses from the real students:

In [21]: model = CTGAN(
   ....:     primary_key='student_id',
   ....: )
   ....: 

In [22]: model.fit(data_pii)

In [23]: new_data_pii = model.sample(200)

In [24]: new_data_pii.head()
Out[24]: 
   student_id                                            address 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           0  4210 Baker Stravenue\nNorth Christopher, WI 90462      F    40.886854  68.071539      Arts    67.089470   Comm&Mgmt             True                 0           60.705155  Mkt&Fin  46.366971           NaN    True 2020-01-16 2020-09-04      6.0
1           1                   Unit 3091 Box 7338\nDPO AP 23696      F    46.628696  92.569276   Science    61.523067   Comm&Mgmt            False                 0           45.885736  Mkt&Fin  54.871093           NaN   False        NaT        NaT      NaN
2           2   1471 Gill Manors Suite 595\nNorth Ryan, CT 04194      F    74.529631  64.216288   Science    49.221393    Sci&Tech            False                 0           63.180622   Mkt&HR  55.591053           NaN    True 2020-07-16 2020-07-06      3.0
3           3   162 Schmidt Harbor Apt. 653\nNew Paula, MI 52655      M    80.530458  69.236880  Commerce    58.604563   Comm&Mgmt            False                 0           78.980979   Mkt&HR  49.765651           NaN    True 2020-01-23 2020-12-13      NaN
4           4  793 Rebecca Isle Apt. 327\nSouth Nicoleport, H...      F    60.506434  74.467124  Commerce    62.781454   Comm&Mgmt             True                 0           60.934057  Mkt&Fin  52.172007  30160.800528    True 2020-03-22 2020-09-07      3.0

More specifically, we can see how all the addresses that have been generated actually come from the original dataset:

In [25]: new_data_pii.address.isin(data_pii.address).sum()
Out[25]: 200

In order to solve this, we can pass an additional argument anonymize_fields to our model when we create the instance. This anonymize_fields argument will need to be a dictionary that contains:

  • The name of the field that we want to anonymize.

  • The category of the field that we want to use when we generate fake values for it.

The list complete list of possible categories can be seen in the Faker Providers page, and it contains a huge list of concepts such as:

  • name

  • address

  • country

  • city

  • ssn

  • credit_card_number

  • credit_card_expire

  • credit_card_security_code

  • email

  • telephone

In this case, since the field is an e-mail address, we will pass a dictionary indicating the category address

In [26]: model = CTGAN(
   ....:     primary_key='student_id',
   ....:     anonymize_fields={
   ....:         'address': 'address'
   ....:     }
   ....: )
   ....: 

In [27]: model.fit(data_pii)

As a result, we can see how the real address values have been replaced by other fake addresses:

In [28]: new_data_pii = model.sample(200)

In [29]: new_data_pii.head()
Out[29]: 
   student_id                                            address 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           0  7770 Mueller Expressway Apt. 608\nEast Valerie...      F    40.127671  70.349477   Science    78.356026      Others            False                 0           90.463599  Mkt&Fin  58.109020            NaN    True 2020-02-01 2020-08-23      NaN
1           1  7546 Rebecca Bridge Suite 038\nNew Zachary, UT...      M    68.775723  66.996412   Science    75.208394   Comm&Mgmt            False                 1           67.194478  Mkt&Fin  69.472567            NaN    True 2020-01-13        NaT      6.0
2           2         78754 Hill Mall\nNew Christopher, WA 61576      F    46.535746  50.574402   Science    84.250139   Comm&Mgmt            False                 0           58.740641   Mkt&HR  68.646390  112284.187131   False        NaT 2020-09-03      NaN
3           3  2670 Dennis Estate Suite 606\nWest Kenneth, CO...      F    43.120468  76.221681  Commerce    71.135291   Comm&Mgmt            False                 0           55.511173  Mkt&Fin  71.048466            NaN    True 2020-01-25 2020-11-27      3.0
4           4         503 Leslie Passage\nJohnsonmouth, IA 88767      F    54.257690  97.344343   Science    75.519884      Others             True                 0           46.282516  Mkt&Fin  58.324091   41612.662088    True 2019-12-22 2020-09-03     12.0

Which means that none of the original addresses can be found in the sampled data:

In [30]: data_pii.address.isin(new_data_pii.address).sum()
Out[30]: 0

Advanced Usage

Now that we have discovered the basics, let’s go over a few more advanced usage examples and see the different arguments that we can pass to our CTGAN Model in order to customize it to our needs.

How to modify the CTGAN Hyperparameters?

A part from the common Tabular Model arguments, CTGAN has a number of additional hyperparameters that control its learning behavior and can impact on the performance of the model, both in terms of quality of the generated data and computational time.

  • epochs and batch_size: these arguments control the number of iterations that the model will perform to optimize its parameters, as well as the number of samples used in each step. Its default values are 300 and 500 respectively, and batch_size needs to always be a value which is multiple of 10.

    These hyperparameters have a very direct effect in time the training process lasts but also on the performance of the data, so for new datasets, you might want to start by setting a low value on both of them to see how long the training process takes on your data and later on increase the number to acceptable values in order to improve the performance.

  • log_frequency: Whether to use log frequency of categorical levels in conditional sampling. It defaults to True. This argument affects how the model processes the frequencies of the categorical values that are used to condition the rest of the values. In some cases, changing it to False could lead to better performance.

  • embedding_dim (int): Size of the random sample passed to the Generator. Defaults to 128.

  • generator_dim (tuple or list of ints): Size of the output samples for each one of the Residuals. A Resiudal Layer will be created for each one of the values provided. Defaults to (256, 256).

  • discriminator_dim (tuple or list of ints): Size of the output samples for each one of the Discriminator Layers. A Linear Layer will be created for each one of the values provided. Defaults to (256, 256).

  • generator_lr (float): Learning rate for the generator. Defaults to 2e-4.

  • generator_decay (float): Generator weight decay for the Adam Optimizer. Defaults to 1e-6.

  • discriminator_lr (float): Learning rate for the discriminator. Defaults to 2e-4.

  • discriminator_decay (float): Discriminator weight decay for the Adam Optimizer. Defaults to 1e-6.

  • discriminator_steps (int): Number of discriminator updates to do for each generator update. From the WGAN paper: https://arxiv.org/abs/1701.07875. WGAN paper default is 5. Default used is 1 to match original CTGAN implementation.

  • verbose: Whether to print fit progress on stdout. Defaults to False.

Warning

Notice that the value that you set on the batch_size argument must always be a multiple of 10!

As an example, we will try to fit the CTGAN model slightly increasing the number of epochs, reducing the batch_size, adding one additional layer to the models involved and using a smaller wright decay.

Before we start, we will evaluate the quality of the previously generated data using the sdv.evaluation.evaluate function

In [31]: from sdv.evaluation import evaluate

In [32]: evaluate(new_data, data)
Out[32]: 0.5230034621786793

Afterwards, we create a new instance of the CTGAN model with the hyperparameter values that we want to use

In [33]: model = CTGAN(
   ....:     primary_key='student_id',
   ....:     epochs=500,
   ....:     batch_size=100,
   ....:     generator_dim=(256, 256, 256),
   ....:     discriminator_dim=(256, 256, 256)
   ....: )
   ....: 

And fit to our data.

In [34]: model.fit(data)

Finally, we are ready to generate new data and evaluate the results.

In [35]: new_data = model.sample(len(data))

In [36]: evaluate(new_data, data)
Out[36]: 0.5190417791199542

As we can see, in this case these modifications changed the obtained results slightly, but they did neither introduce dramatic changes in the performance.

How do I specify constraints?

If you look closely at the data you may notice that some properties were not completely captured by the model. For example, you may have seen that sometimes the model produces an experience_years number greater than 0 while also indicating that work_experience is False. These type of properties are what we call Constraints and can also be handled using SDV. For further details about them please visit the Handling Constraints guide.

Can I evaluate the Synthetic Data?

A very common question when someone starts using SDV to generate synthetic data is: “How good is the data that I just generated?”

In order to answer this question, SDV has a collection of metrics and tools that allow you to compare the real that you provided and the synthetic data that you generated using SDV or any other tool.

You can read more about this in the Synthetic Data Evaluation guide.