An open source project from Data to AI Lab at MIT.
Automated generative modeling and sampling
- Free software: MIT license
- Documentation: https://HDI-Project.github.io/SDV
The goal of the Synthetic Data Vault (SDV) is to allow data scientists to navigate, model and
sample relational databases. The main access point of the library is the class SDV
, that wraps
the functionality of the three core classes: the DataNavigator
, the Modeler
and the Sampler
.
Using these classes, users can get easy access to information about the relational database, create generative models for tables in the database and sample rows from these models to produce synthetic data.
The easiest way to install SDV is using pip
pip install sdv
You can also clone the repository and install it from sources
git clone [email protected]:HDI-Project/SDV.git
cd SDV
pip install -e .
Below there is a short example about how to use SDV to model and sample a dataset composed of relational tables.
The easiest way to use SDV is using the SDV class from the root of the package:
>>> from sdv import SDV
>>> data_vault = SDV('tests/data/meta.json')
>>> data_vault.fit()
>>> samples = data_vault.sample_all()
>>> for dataset in samples:
... print(samples[dataset].head(3), '\n')
CUSTOMER_ID CUST_POSTAL_CODE PHONE_NUMBER1 CREDIT_LIMIT COUNTRY
0 0 61026.0 5.410825e+09 1017.0 FRANCE
ORDER_ID CUSTOMER_ID ORDER_TOTAL
0 0 0 1251.0
1 1 0 1691.0
2 2 0 1126.0
ORDER_ITEM_ID ORDER_ID PRODUCT_ID UNIT_PRICE QUANTITY
0 0 0 9.0 20.0 0.0
1 1 0 8.0 79.0 3.0
2 2 0 8.0 66.0 1.0
With this, we will be able to generate sintetic samples of data. The only argument we pass to SDV
is a path to a JSON file containing the information of the different tables, their fields and
relations. Further explanation of how to generate this file can be found on the docs.
After instantiating the class, we call to the fit()
method in order to transform and model the
data, and after that we are ready to sample rows, tables or the whole database.
The overall flow of SDV is as follows: the DataNavigator extracts relevant information from the
dataset, as well as applies desired transformations. This class is then passed into a Modeler
,
which uses the information in the DataNavigator
to create generative models of the tables.
Finally, a Modeler
can be given to a Sampler
to actual sample rows of synthetic data.
The DataNavigator
can be used to extract useful information about a dataset, such as the
relationships between tables. It can also be used to apply transformations. Here we will use it to
load the test data from the CSV files and apply some transformations to it.
First, we will instantiate the CSVDataLoader
class, that will load the data and prepare it to use it with DataNavigator
.
To create an instance of the CSVDataLoader
class, the filepath to the meta.json file must be provided.
>>> from sdv import CSVDataLoader
>>> data_loader = CSVDataLoader('tests/data/meta.json')
The load_data()
function can then be used to create an instance of a DataNavigator
.
>>> data_navigator = data_loader.load_data()
The DataNavigator
stores the data as a dictionary mapping the table names to a tuple of the data
itself (represented as a pandas.Dataframe
) and the meta information for that table. You can access
the data using the following command:
>>> customer_table = data_navigator.tables['DEMO_CUSTOMERS']
>>> customer_data = customer_table.data
>>> customer_data.head(3).T
0 1 2
CUSTOMER_ID 50 4 97338810
CUST_POSTAL_CODE 11371 63145 6096
PHONE_NUMBER1 6175553295 8605551835 7035552143
CREDIT_LIMIT 1000 500 1000
COUNTRY UK US CANADA
>>> customers_meta = customer_table.meta
>>> customers_meta.keys()
dict_keys(['fields', 'headers', 'name', 'path', 'primary_key', 'use'])
>>> customers_meta['fields']
{'CUSTOMER_ID': {'name': 'CUSTOMER_ID',
'subtype': 'integer',
'type': 'number',
'uniques': 0,
'regex': '^[0-9]{10}$'},
'CUST_POSTAL_CODE': {'name': 'CUST_POSTAL_CODE',
'subtype': 'integer',
'type': 'number',
'uniques': 0},
'PHONE_NUMBER1': {'name': 'PHONE_NUMBER1',
'subtype': 'integer',
'type': 'number',
'uniques': 0},
'CREDIT_LIMIT': {'name': 'CREDIT_LIMIT',
'subtype': 'integer',
'type': 'number',
'uniques': 0},
'COUNTRY': {'name': 'COUNTRY', 'type': 'categorical', 'uniques': 0}}
You can also use the data navigator to get parents or children of a table.
>>> data_navigator.get_parents('DEMO_ORDERS')
{'DEMO_CUSTOMERS'}
>>> data_navigator.get_children('DEMO_CUSTOMERS')
{'DEMO_ORDERS'}
Finally, we can use the transform_data()
function to apply transformations from the
RDT library to our data. If no transformations are provided,
the function will convert all categorical types and datetime types to numeric values by default.
It will return a dictionary mapping the table name to the transformed data represented as a
pandas.Dataframe
.
>>> transformed_data = data_navigator.transform_data()
>>> transformed_data['DEMO_CUSTOMERS'].head(3).T
0 1 2
CUSTOMER_ID 5.000000e+01 4.000000e+00 9.733881e+07
CUST_POSTAL_CODE 1.137100e+04 6.314500e+04 6.096000e+03
PHONE_NUMBER1 6.175553e+09 8.605552e+09 7.035552e+09
CREDIT_LIMIT 1.000000e+03 5.000000e+02 1.000000e+03
COUNTRY 5.617796e-01 8.718027e-01 5.492714e-02
The Modeler
can be used to recursively model the data. This is important because the tables in
the data have relationships between them, that should also be modeled in order to have reliable
sampling. Let's look at the test data for example. There are three tables in this data set:
DEMO_CUSTOMERS
, DEMO_ORDERS
and DEMO_ORDER_ITEMS
.
The DEMO_ORDERS
table has a field labelled CUSTOMER_ID
, that references the "id" field
of the DEMO_CUSTOMERS
table. SDV wants to model not only the data, but these relationships as
well. The Modeler class is responsible for carrying out this task.
To do so, first, import from the Modeler and create an instance of the class. The Modeler must be provided the DataNavigator and the type of model to use. If no model type is provided, it will use a copulas.multivariate.Gaussian Copula by default. Note that in order for the modeler to work, the DataNavigator must have already transformed its data.
>>> from sdv import Modeler
>>> modeler = Modeler(data_navigator)
Then you can model the entire database. The modeler will store models for every table in the dataset.
>>> modeler.model_database()
The models that were created for each table can be accessed using the following command:
>>> customers_model = modeler.models['DEMO_CUSTOMERS']
>>> print(customers_model)
CUSTOMER_ID
==============
Distribution Type: Gaussian
Variable name: CUSTOMER_ID
Mean: 22198555.57142857
Standard deviation: 36178958.000449404
CUST_POSTAL_CODE
==============
Distribution Type: Gaussian
Variable name: CUST_POSTAL_CODE
Mean: 34062.71428571428
Standard deviation: 25473.85661931119
PHONE_NUMBER1
==============
Distribution Type: Gaussian
Variable name: PHONE_NUMBER1
Mean: 6464124184.428572
Standard deviation: 1272684276.6679976
...
The output above shows the parameters that got stored for every column in the users table.
The modeler can also be saved to a file using the save()
method. This will save a pickle file
on the specified path.
>>> modeler.save('models/demo_model.pkl')
If you have stored a model in a previous session using the command above, you can load the model
using the load()
method:
>>> modeler = Modeler.load('models/demo_model.pkl')
The Sampler
takes in a Modeler
and DataNavigator
. Using the mdels created in the last step,
the Sampler
can recursively move through the tables in the dataset, and sample synthetic data.
It can be used to sample rows from specified tables, sample an entire table at once or sample the
whole database.
Let's do an example with out dataset. First import the Sampler and create an instance of the class.
>>> from sdv import Sampler
>>> sampler = Sampler(data_navigator, modeler)
To sample from a row, use the command sample_rows()
. Note that before sampling from a child table, one of
its parent tables must be sampled from.
>>> sampler.sample_rows('DEMO_CUSTOMERS', 1).T
0
CUSTOMER_ID 0
CUST_POSTAL_CODE 44462
PHONE_NUMBER1 7.45576e+09
CREDIT_LIMIT 976
COUNTRY US
To sample a whole table use sample_table()
. This will create as many rows as in the original
database.
>>> sampler.sample_table('DEMO_CUSTOMERS')
CUSTOMER_ID CUST_POSTAL_CODE PHONE_NUMBER1 CREDIT_LIMIT COUNTRY
0 0 27937.0 8.095336e+09 1029.0 CANADA
1 1 18183.0 2.761015e+09 891.0 CANADA
2 2 16402.0 4.956798e+09 1313.0 SPAIN
3 3 7116.0 8.072395e+09 1124.0 FRANCE
4 4 368.0 4.330203e+09 1186.0 FRANCE
5 5 64304.0 6.256936e+09 1113.0 US
6 6 94698.0 8.271224e+09 1086.0 CANADA
Finally, the entire database can be sampled using sample_all(num_rows)
. The num_rows
parameter
specifies how many child rows to create per parent row. This function returns a dictionary mapping
table names to the generated dataframes.
>>> samples = sampler.sample_all()
>>> for dataset in samples:
... print(samples[dataset].head(3), '\n')
CUSTOMER_ID CUST_POSTAL_CODE PHONE_NUMBER1 CREDIT_LIMIT COUNTRY
0 0 46038.0 7.779893e+09 673.0 UK
1 1 21063.0 6.511387e+09 808.0 SPAIN
2 2 24494.0 5.703648e+09 757.0 SPAIN
ORDER_ID CUSTOMER_ID ORDER_TOTAL
0 0 0 1520.0
1 1 0 1217.0
2 2 0 1375.0
ORDER_ITEM_ID ORDER_ID PRODUCT_ID UNIT_PRICE QUANTITY
0 0 0 17.0 94.0 3.0
1 1 0 14.0 44.0 3.0
2 2 0 20.0 78.0 3.0