Systems and methods for generating synthetic tabular data for machine learning and other applications

1 . A system comprising:

a memory component; and

one or more processing devices coupled to the memory component, the one or more processing devices to perform operations comprising:

receiving a set of tabular data records, each tabular data record of the set of tabular data records comprising a plurality of features;

training a first machine learning model using the set of tabular data records to learn one or more correlations between the plurality of features, the first machine learning model comprising a variational autoencoder; and

training a second machine learning model, using the first machine learning model, to generate a set of synthetic tabular data records based at least on the one or more correlations between the plurality of features, the second machine learning model comprising a generative adversarial network that includes a generator model and a critic model, the variational autoencoder coupled between the generator model and the critic model.

2 . The system of claim 1 , wherein the first machine learning model comprises the variational autoencoder, the operations further comprising:

training the variational autoencoder to learn the one or more correlations between the plurality of features by iteratively:

producing a noised tabular data record sample by introducing a noise to a first tabular data record of the set of tabular data records, wherein the noise masks a feature of the first tabular data record;

generating a denoised tabular data record sample based on the noised tabular data record sample, wherein the denoised tabular data record sample comprises a prediction of the feature masked by the noise; and

adjusting the variational autoencoder based on a reconstruction loss computed from a difference between the denoised tabular data record sample and the first tabular data record of the set of tabular data records.

3 . The system of claim 1 , wherein the first machine learning model comprises the variational autoencoder and the second machine learning model comprises a generative adversarial network that includes a generator model and a critic model, the operations further comprising:

training the generative adversarial network using the variational autoencoder to generate the set of synthetic tabular data records by iteratively:

generating an initial synthetic tabular data record sample with the generator model, the initial synthetic tabular data record sample comprising the plurality of features;

producing a noised synthetic tabular data record sample by introducing a noise to the initial synthetic tabular data record sample, wherein the noise masks a feature of the initial synthetic tabular data record sample;

generating a denoised synthetic tabular data record sample using the variational autoencoder based on the noised synthetic tabular data record sample, wherein the denoised synthetic tabular data record sample comprises a prediction of the feature of the initial synthetic tabular data record sample masked by the noise; and

adjusting one or both of the generator model and the critic model based on a reconstruction loss computed from a difference between the denoised synthetic tabular data record sample and the initial synthetic tabular data record sample.

4 . The system of claim 3 , the operations further comprising:

adjusting the generator model based on a condition vector; and

generating the initial synthetic tabular data record sample by conditioning a first feature of the plurality of features based at least on the condition vector.

5 . The system of claim 4 , the operations further comprising:

computing the condition vector based at least in part on a distribution of the plurality of features within the set of tabular data records.

6 . The system of claim 3 , the operations further comprising:

storing the denoised synthetic tabular data record sample to a set of denoised synthetic tabular data records; and

determining the reconstruction loss at least in part based on a Pearson coefficient matrix computed based on the set of tabular data records and the set of denoised synthetic tabular data records.

7 . The system of claim 3 , the operations further comprising:

storing the denoised synthetic tabular data record sample to a set of denoised synthetic tabular data records;

determining a first distribution characteristic based on the set of tabular data records and determining a second distribution characteristic based on the set of denoised synthetic tabular data records, wherein the critic model computes a realness score using the first distribution characteristic and the second distribution characteristic; and

adjusting one or both of the generator model and the critic model based on a generator loss computed from the realness score.

8 . The system of claim 1 , wherein the second machine learning model comprises a generative adversarial network that includes a generator model and a critic model, the operations further comprising:

generating the set of synthetic tabular data records with the generator model; and

training a third machine model using the set of synthetic tabular data records.

9 . A method comprising:

producing a variational autoencoder by training an encoder-decoder machine learning model to learn one or more correlations between a plurality of features of a set of tabular data records based on noised tabular data record samples derived from the set of tabular data records;

producing a generative adversarial network that comprises a generator model, a critic model, and the variational autoencoder, wherein the variational autoencoder is coupled between the generator model and the critic model; and

training the generator model to generate a set of synthetic tabular data records based at least on the one or more correlations between the plurality of features using a first loss computed from an output of the variational autoencoder and a second loss computed from an output of the critic model.

10 . The method of claim 9 , further comprising:

training the encoder-decoder machine learning model to learn the one or more correlations between the plurality of features by iteratively:

producing a noised tabular data record sample by introducing a noise to a first tabular data record of the set of tabular data records, wherein the noise masks a feature of the first tabular data record;

generating a denoised tabular data record sample based on the noised tabular data record sample, wherein the denoised tabular data record sample comprises a prediction of the feature masked by the noise; and

adjusting the encoder-decoder machine learning model based on a reconstruction loss computed from a difference between the denoised tabular data record sample and the first tabular data record of the set of tabular data records.

11 . The method of claim 10 , further comprising:

designating a first feature of the plurality of features as a protected feature; and

wherein the first feature is not masked by the noise when training the encoder-decoder machine learning model.

12 . The method of claim 9 , further comprising:

training the generative adversarial network using the variational autoencoder to generate the set of synthetic tabular data records by iteratively:

generating an initial synthetic tabular data record sample with the generator model, the initial synthetic tabular data record sample comprising the plurality of features;

producing a noised synthetic tabular data record sample by introducing a noise to the initial synthetic tabular data record sample, wherein the noise masks a feature of the initial synthetic tabular data record sample;

generating a denoised synthetic tabular data record sample using the variational autoencoder based on the noised synthetic tabular data record sample, wherein the denoised synthetic tabular data record sample comprises a prediction of the feature of the initial synthetic tabular data record sample masked by the noise; and

adjusting one or both of the generator model and the critic model based on a reconstruction loss computed from a difference between the denoised synthetic tabular data record sample and the initial synthetic tabular data record sample.

13 . The method of claim 12 , further comprising:

storing the denoised synthetic tabular data record sample to a set of denoised synthetic tabular data records; and

determining the reconstruction loss at least in part based on a Pearson coefficient matrix computed based on the set of tabular data records and the set of denoised synthetic tabular data records.

14 . The method of claim 12 , further comprising:

storing the denoised synthetic tabular data record sample to a set of denoised synthetic tabular data records;

determining a first distribution characteristic based on the set of tabular data records and determining a second distribution characteristic based on the set of denoised synthetic tabular data records, wherein the critic model computes a realness score using the first distribution characteristic and the second distribution characteristic; and

adjusting one or both of the generator model and the critic model based on a generator loss computed from the realness score.

15 . The method of claim 12 , further comprising:

adjusting the generator model based on a condition vector; and

generating the initial synthetic tabular data record sample by conditioning a first feature of the plurality of features based at least on the condition vector.

16 . The method of claim 15 , further comprising:

computing the condition vector based at least in part on a distribution of the plurality of features within the set of tabular data records.

17 . A system comprising:

a memory component; and

one or more processing devices coupled to the memory component, the one or more processing devices to perform operations comprising:

executing a generator model, wherein the generator model was trained using a generative adversarial network architecture comprising the generator model, a critic model, and a denoising variational autoencoder coupled between the generator model and the critic model;

generating, using the generator model, a first set of synthetic tabular data records, wherein respective synthetic tabular data records individually comprise a plurality of features and having a first correlation between features of the plurality of features, the first correlation based at least in part on a second correlation between features of a second set of tabular data records, the generator model further trained to learn the second correlation based on a loss determined using the denoising variational autoencoder; and

storing the first set of synthetic tabular data records to a data store.

18 . The system of claim 17 , the operations further comprising:

executing a machine learning model training application to train a second machine learning model based at least on the first set of synthetic tabular data records.

19 . The system of claim 17 , the operations further comprising:

adjusting the generator model using a condition vector; and

generating the first set of synthetic tabular data records by conditioning a feature of the plurality of features based at least on the condition vector.