datgan 2.1.2

Generative adversarial network with integrated expert knowledge for synthesizing tabular data

DATGAN

Directed Acyclic Tabular GAN (DATGAN) for integrating expert knowledge in synthetic tabular data generation.

• License: GPLv3
• Development Status: Alpha
• Homepage: https://github.com/glederrey/DATGAN

Overview

The DATGAN is a synthesizer for tabular data. It uses LSTM cells to generate synthetic data for continuous and categorical variable types. In addition, a Directed Acyclic Graph (DAG) can be provided to represent the structure between the variables and help the model to perform better. This model integrates two types of conditionality: rejection by sampling and conditional inputs.

Requirements

The current version (v2.1.1) of the DATGAN works with Python 3.9 and Tensorflow 2. Previous versions of Python have not been tested and are thus blocked. We, thus, recommend the user to set up a virtualenv. An installation guid is provided to help with the installation of the library.

Installation

We provide a complete installation guide using conda and setting up a virtualenv. Please follow this guide to properly set up everything and make sure that you can use the DATGAN as intended.

Testing the DATGAN

Once you have followed the installation guid, you can clone this repository and use the notebooks provided in the folder example to train the DATGAN and use the evaluation metrics provided in this repository.

Data Format

Input

The DATGAN uses tabular data loaded with the pandas library. This table must follow these rules:

• has no missing values
• has columns of types int, float, str or bool.
• each column contains data of only one type.

NOTE: It is important to identify which columns are considered continuous and which are considered categorical. A general rule of thumb is that any discrete distributions with a large number of categories should be considered continuous, and then rounded.

Output

The output of the DATGAN is a table of synthetic data with the same columns as the input table and as many rows as requested.

Tutorial

In this short tutorial we will guide you through a series of steps that will help you to get started with the most basic usage of DATGAN in order to generate samples from a given dataset.

While all the methods are explained below, you can use the helper built-in function to get a reminder of the different functions that are used in this library. Just run the following lines:

import datgan
help(datgan)

1. Load the data and provide info about it

The first step is to load the data which we will use to fit the DATGAN. In the example, we provide a demo dataset, the CMAP dataset. It is a travel survey dataset for the city of Chicago. The dataset has already been cleaned compared to the original dataset. It contains 15 columns and 8'929 rows.

You can load this dataset using pandas. You also need to provide the type of data for each columns. For the moment, only two possibilities are implemented: continuous and categorical. For continuous columns, you can pass more information to the model such as:

• discrete [mandatory]: Boolean value to indicate if the synthetic value has to be rounded when sampling
• bounds [optional]: Values of the bounds. While sampling synthetic data, all values outside the bounds will be discarded.
• enforce_bounds [optional]: Instead of removing synthetic data generated outside the bounds, we will clip the values to the bounds. (This can be useful in the case of a mixed distribution with a peak of values close to the lower or the upper bound.)
• apply_func [optional]: You can provide a lambda function that will be applied before the encoding step and when sampling the final values. This can help to train models on distributions that are more difficult to be represented by GMMs, e.g. exponential distributions.

import pandas as pd
import numpy as np

df = pd.read_csv('./data/CMAP.csv', index_col=False)

data_info = {
    'distance': {
        'type': 'continuous',
        'bounds': [0.0, np.infty],
        'discrete': False,
        'apply_func': (lambda x: np.log(x+1)),
    },
    'age': {
        'type': 'continuous',
        'bounds': [0, 100],
        'enforce_bounds': True,
        'discrete': True
    },
    'departure_time': {
        'type': 'continuous',
        'bounds': [0, 23.999],
        'discrete': False
    }
}

# For the categorical columns, we can simply add them using a for loop
for c in df.columns:
    if c not in data_info.keys():
        data_info[c] = {'type': 'categorical'}

WARNING: While using a lambda function in the apply_func parameters, we are using the pynverse library to compute the inverse of the given function. Therefore, with more complex functions, it is possible to get Warnings while sampling the final synthetic data. This is a normal behaviour when pynverse is having trouble computing the inverse of some values.

2. Create a DAG

The second steps consists in creating the DAG for the DATGAN. The DAG is created using the library networkx from Python. The only constraint while creating the DAG is that it should not contain any cycle (by definition). In the article, we give more details how to create the DAG. However, this process can still be cumbersome and requires trials and errors to obtain a good DAG. (We highly recommend using a "visualization" tool to create the DAG.)

If you do not have any idea how to create your DAG, it is possible to not provide any DAG to the model. However, in this case, the model will define a linear DAG, i.e. each variable in the dataset is linked to the next one following the order of the columns. It can be useful to quickly test the model. However, it will reduce the performance of the model as shown in the article.

Example of a DAG for the CMAP dataset:

import networkx as nx

graph = nx.DiGraph()
graph.add_edges_from([
    ("age", "license"),
    ("age", "education_level"),
    ("gender", "work_status"),
    ("education_level", "work_status"),
    ("education_level", "hh_income"),
    ("work_status", "hh_income"),
    ("hh_income", "hh_descr"),
    ("hh_income", "hh_size"),
    ("hh_size", "hh_vehicles"),
    ("hh_size", "hh_bikes"),
    ("work_status", "trip_purpose"),
    ("trip_purpose", "departure_time"),
    ("trip_purpose", "distance"),
    ("travel_dow", "choice"),
    ("distance", "choice"),
    ("departure_time", "choice"),
    ("hh_vehicles", "choice"),
    ("hh_bikes", "choice"),
    ("license", "choice"),
    ("education_level", "hh_size"),
    ("work_status", "hh_descr"),
    ("work_status", "hh_size"),
    ("hh_income", "hh_bikes"),
    ("hh_income", "hh_vehicles"),
    ("trip_purpose", "choice")
])

We also implemented a function named advise that computes Pearson and Spearman correlations and then suggests some links that could be added to the DAG. The links suggested are the one with the highest correlations.

import datgan

datgan.advise(df, graph, plot_graphs=True)

Results (without the graphs) should look like this:

You might want to add the following edges in your DAG (direction not given here):
  - license <-> education_level
  - license <-> work_status
  - hh_vehicles <-> hh_descr
  - hh_bikes <-> age
  - hh_size <-> age

3. Create a DATGAN instance

The next step is to import DATGAN and create an instance of the model. There are no required parameters for the model. However, we advise you to set up the basic parameters such as the output folder (output), batch size (batch_size), and the number of epochs (num_epochs).

output_folder = './output/'
batch_size = 1116

from datgan import DATGAN

datgan = DATGAN(output=output_folder, batch_size=batch_size, num_epochs=1000)

NOTE: Setting up a suitable batch size is really important. A batch size too big will make the model crash due to memory error while one that is too small will make the model slower to train. Trials and errors are required depending on your hardware. In addition, it is good to find a batch size such that len(df) % batch_size is as small as possible since the last batch of data is dropped if it is smaller than the batch size.

4. Preprocess the data (optional)

The fourth step consists in preprocessing the data. This step is optional since it is automatically done in the next step if skipped right now. We propose to do the preprocessing in advance because it usually takes a bit of time. And if you want to try multiple parameters with the DATGAN, you do not have to preprocess the data every time. Therefore, it is possible to do it before fitting the model and saving it somewhere.

datgan.preprocess(df, data_info, preprocessed_data_path='./encoded_data')

NOTE 1: If you decide to change any values in the metadata dictionary data_info, you will have to preprocess the data again. However, since the preprocessing of the continuous variable is random, you will not be able to sample a model trained on a previous preprocessed dataset with the current one. Therefore, we advise to not delete the previous preprocessed dataset.

NOTE 2: In the folder preprocessed_data_path/continuous, you will find PNG files of the continuous distributions. These pictures can be used to make sure that the Variational Gaussian Mixtures (VGM) were correctly trained for each continuous distributions. It can be especially useful if the user is trying to use a lambda function.

5. Fit the model

Once you have a DATGAN instance, you can call the method fit and passing the following parameters:

• data: the original DataFrame
• graph: the networkx DAG
• continuous_columns: the list of continuous columns
• preprocessed_data_path: the path to the preprocessed data if done in Step 4 or the path where to save them.

datgan.fit(df, data_info, graph, preprocessed_data_path='./encoded_data')

NOTE: Depending on the level of verbose, the model will show different type of information. The standard level is 1, and it will display a progress bar for the whole training process. With level 2, the model will display information at each epoch. With level 0, the model will not print any information.

6. Sample new data

Once the model has been fitted, you can generate new synthetic data by calling the function sample. You have to provide the desired number of samples.

samples = datgan.sample(len(df))
samples.to_csv('./data/synthetic.csv', index=False)

7. Save and load a model (optional)

In the steps above we saw that the fitting process can take a lot of time, so we probably would like to avoid having to fit every we want to generate samples. We advise the use to save checkpoints of the model while it is training. However, if you do not want to do that, the model will always save the latest checkpoint once it has finished training. You can, thus, load it at any time afterwards.

In order to load the model, you can simply call the function load with the parameters used while fitting the model. In order to save memory, we only save the parameters of the Generator and Discriminator. Therefore, more information is required to load the model.

new_datgan = datgan.load(df, graph, preprocessed_data_path='./encoded_data')

NOTE: This can be replaced by using the function fit once again. Since the model has already been trained, it will just load it.

Conditionality

Conditionality can be used in two different ways. The first one is done after training the DATGAN model. It consists in providing a dictionary of desired values and pass it to the sampling method. the DATGAN model uses rejection sampling to provide the desired output. The second method consists in training the DATGAN using conditional inputs. In order to use the second conditionality, the model has to be trained again.

Rejection sampling

In the sample function of the DATGAN, you must provide a dictionary as follows:

• categorical variables: string or list of string for the desired categories.
• continuous variables: lambda function returning a boolean value.

For example, if you want to only generate synthetic data about men under 30 years old, you can provide the following dictionary to the sample function.

cond_dict = {'age': lambda x: x<30, 'gender': '0'}

samples = datgan.sample(len(df), cond_dict=cond_dict)
samples.to_csv('./data/synthetic.csv', index=False)

WARNING: Since the DATGAN is using rejection sampling to provide the desired values, you cannot provide a conditional dictionary that is too restrictive. This will significantly slow the sampling process, and it can even make it fail. In some very specific case, you can set the parameter timeout to False if you do not want the sampling process to be stopped.

Conditional inputs

The conditional inputs are columns present in the original dataset. In order to train the DATGAN model, you need to provide the list of columns you want to consider as conditional inputs. For example, if you want the age, the gender, and the trip purpose to be conditional inputs, you can initialize the DATGAN model this way:

datgan = DATGAN(output=output_folder, 
                batch_size=batch_size, 
                num_epochs=1000,
                conditional_inputs=['age', 'gender', 'trip_purpose'])

If the model is initialized in such way, it will automatically transform the DAG such that each of the conditional variables are considered as source nodes in the DAG. It will then train the Generator using the values for these variables as inputs. Finally, in order to sample the synthetic data, you need to provide a pandas.DataFrame or a dict with columns/keys corresponding to the conditional inputs defined in the initialization. You do not need to care about the size of the DataFrame since the model automatically sample values from it. You can, thus, provide any values you want (subset of original dataset, unique values, or values generated by another model on another dataset.)

In addition, the rejection sampling conditionality can be used alongside the conditional inputs. You just have to provide the dictionary of conditionals alongside the new input data.

Model parameters

If you want to change the default behavior of DATGAN, such as using different batch size or the total number of epochs, you can do so by passing different arguments when creating the DATGAN instance or using its two main functions.

Initialize the model (DATGAN.__init__)

Name	Type	Default	Explanation
loss_function	str	None	Name of the loss function to be used. If not specified, it will choose between 'WGAN' and 'WGGP' depending on the ratio of continuous and categorical columns. Only accepts the values 'SGAN', 'WGAN', and 'WGGP'.
label_smoothing	str	'TS'	Type of label smoothing. Only accepts the values 'TS', 'OS', and 'NO'.
output	str	'./output'	Path to store the model and its artifacts.
gpu	int	None	Model will automatically try to use GPU if tensorflow can use CUDA. However, this parameter allows you to choose which GPU you want to use.
num_epochs	int	100	Number of epochs to use during training.
batch_size	int	500	Size of the batch to feed the model at each step.
save_checkpoints	bool	True	Whether to store checkpoints of the model after each training epoch.
restore_session	bool	True	Whether continue training from the last checkpoint.
learning_rate	float	None	Learning rate. If set to None, the value will be set according to the chosen loss function.
g_period	int	None	Every g_period steps, train the generator once. (Used to train the discriminator more than the generator) By default, it will choose values according the chosen loss function.
l2_reg	bool	None	Tell the model to use L2 regularization while training both NNs. By default, it only applies the L2 regularization when using the SGAN loss function.
z_dim	int	200	Dimension of the noise vector used as an input to the generator.
num_gen_rnn	int	100	Size of the hidden units in the LSTM cell.
num_gen_hidden	int	50	Size of the hidden layer used on the output of the generator to act as a convolution.
num_dis_layers	int	1	Number of layers for the discriminator.
num_dis_hidden	int	100	Size of the hidden layers in the discriminator.
noise	float	0.2	Upper bound to the gaussian noise added to with the label smoothing. (only used if label_smoothing is set to 'TS' or 'OS')
conditional_inputs	list[str]	None	List of variables in the dataset that are used as conditional inputs to the model.
verbose	int	1	Level of verbose. 0 means no print, 1 means that some details will be printed, 2 is mostly used for debugging purpose.

Training the model (DATGAN.fit)

Name	Type	Default	Explanation
data	pandas.DataFram		Original dataset
metadata	dict	None	Dictionary containing information about the data in the DataFrame.
dag	networkx.DiGraph	None	Directed Acyclic Graph representing the relations between the variables. If no dag is provided, the algorithm will create a linear DAG.
preprocess_data_path	str	None	Path to an existing preprocessor. If None is given, the model will preprocess the data and save it under self.output + '/encoded_data'.

Sampling synthetic data (DATGAN.sample)

Name	Type	Default	Explanation
num_samples	int		Number of rows in the generated synthetic dataset.
inputs	dict or pandas.DataFrame	None	Optional input data. Must be provided if the model was training with some conditional inputs in the variable conditional_inputs.
cond_dict	dict	None	Dictionary for the rejection sampling conditionality.
sampling	str	SS	Sampling strategy. Only accepts the following values: 'SS', 'SA', 'AS', and 'AA'.
timeout	bool	True	Timeout to stop sampling if the model cannot generated the required data in cond_dict.

When sampling the synthetic data, you can choose between multiple sampling strategies. S means we are using simulation to sample the data, A means that we are using argmax. The first letter corresponds to continuous variables and the second to categorical variables. Therefore, SA means we're using simulation for continuous variables and argmax for categorical variables.

Tips and tricks

While the DATGAN model will automatically choose the parameters of the model if none are provided, we highly recommend the user to "play" with them. The most important ones are the following:

• loss_function Generally, the WGAN loss function works better on datasets with more categorical columns than continuous. It is the contrary for the WGGP loss. The SGAN loss seems to perform a bit less good than the other two in the specific cases. However, it seems to perform ok in any cases.
• g_period This parameter is especially important when using either the WGAN or the WGGP loss. Generally, the latter requires a lot more training of the discriminator than the previous. However, it might be interesting to test different values to see which one leads to the best results.
• l2_reg The L2 regularization is "mandatory" for the SGAN loss and "forbidden" for the WGAN loss. The user can test these other configurations, but it will lead to worse results. However, for the WGGP loss, there are no specific rules whether to apply it or not. Therefore, it might be interesting to test this parameter as well.
• learning_rate It has been fixed depending on the loss function. However, as for any optimization problem, playing with the learning rate to find the optimal value is always important.

Contributing

If you want to participate in the development of this model, you can write a Github issue and give us your suggestions or directly prepare a Pull Request. (However, it would be better to first discuss the content of the PR.)

Acknowledgements

We would like to thank the authors of the TGAN model, the authors of the CTGAN model, and the authors of the CTAB-GAN model. These three models have greatly inspired the DATGAN model.

Citation

If you use DATGAN or its evaluation metrics for any of your projects, please cite the following work:

"Lederrey G., Hillel T., Bierlaire M., DATGAN: Integrating Expert Knowledge into Deep Learning for Synthetic Tabular Data, arXiv preprint arXiv:2203.03489 (2022)"

The original code for this article can be found in this Github repository: https://github.com/glederrey/SynthPop.