synthcity.plugins.privacy.plugin_privbayes module

Reference: PrivBayes: Private Data Release via Bayesian Networks. (2017), Zhang J, Cormode G, Procopiuc CM, Srivastava D, Xiao X.

class PrivBayes(epsilon: float = 1.0, K: int = 0, n_bins: int = 100, mi_thresh: float = 0.01, target_usefulness: int = 5)

Bases: synthcity.plugins.core.serializable.Serializable

PrivBayes is a differentially private method for releasing high-dimensional data.

Given a dataset D, PrivBayes first constructs a Bayesian network N , which

  1. provides a succinct model of the correlations among the attributes in D

  2. allows us to approximate the distribution of data in D using a set P of lowdimensional marginals of D.

After that, PrivBayes injects noise into each marginal in P to ensure differential privacy, and then uses the noisy marginals and the Bayesian network to construct an approximation of the data distribution in D. Finally, PrivBayes samples tuples from the approximate distribution to construct a synthetic dataset, and then releases the synthetic data.

display_network() → None
fit(data: pandas.core.frame.DataFrame) → Any
static load(buff: bytes) → Any
static load_dict(representation: dict) → Any
mutual_info_score(data: pandas.core.frame.DataFrame, parents: List[str], candidate: str) → float

Cluster the source columns, and compute the mutual information between the target and the clusters.

sample(count: int) → pandas.core.frame.DataFrame
save() → bytes
save_dict() → dict
save_to_file(path: pathlib.Path) → bytes
static version() → str

API version

class PrivBayesPlugin(epsilon: float = 1.0, K: int = 0, n_bins: int = 100, mi_thresh: float = 0.01, target_usefulness: int = 5, random_state: int = 0, workspace: pathlib.Path = PosixPath('workspace'), compress_dataset: bool = False, sampling_patience: int = 500, **kwargs: Any)

Bases: synthcity.plugins.core.plugin.Plugin

Inheritance diagram of synthcity.plugins.privacy.plugin_privbayes.PrivBayesPlugin

PrivBayes algorithm.

Args:
epsilon: float

Differential privacy parameter

K:

Maximum number of parents for a node

n_bins: int

Number of bins for encoding the features

mi_thresh: int

Mutual information lower threshold. If the current score is lower, the [] parents are used.

target_usefulness: int

Def 4.7 in the paper: A noisy distribution is θ-useful if the ratio of average scale of

information to average scale of noise is no less than θ. 5-useful is the recommended value.
random_state: int

Random seed

# Core Plugin arguments workspace: Path.

Optional Path for caching intermediary results.

compress_dataset: bool. Default = False.

Drop redundant features before training the generator.

sampling_patience: int.

Max inference iterations to wait for the generated data to match the training schema.

Example

>>> from sklearn.datasets import load_iris
>>> from synthcity.plugins import Plugins
>>>
>>> X, y = load_iris(as_frame = True, return_X_y = True)
>>> X["target"] = y
>>>
>>> plugin = Plugins().get("privbayes")
>>> plugin.fit(X)
>>>
>>> plugin.generate(50)
class Config

Bases: object

arbitrary_types_allowed = True
validate_assignment = True
fit(X: Union[synthcity.plugins.core.dataloader.DataLoader, pandas.core.frame.DataFrame], *args: Any, **kwargs: Any) → Any

Training method the synthetic data plugin.

Parameters

  • X – DataLoader. The reference dataset.

  • cond –

    Optional, Union[pd.DataFrame, pd.Series, np.ndarray] Optional Training Conditional. The training conditional can be used to control to output of some models, like GANs or VAEs. The content can be anything, as long as it maps to the training dataset X. Usage example:

    >>> from sklearn.datasets import load_iris
>>> from synthcity.plugins.core.dataloader import GenericDataLoader
>>> from synthcity.plugins.core.constraints import Constraints
>>>
>>> # Load in `test_plugin` the generative model of choice
>>> # ....
>>>
>>> X, y = load_iris(as_frame=True, return_X_y=True)
>>> X["target"] = y
>>>
>>> X = GenericDataLoader(X)
>>> test_plugin.fit(X, cond=y)
>>>
>>> count = 10
>>> X_gen = test_plugin.generate(count, cond=np.ones(count))
>>>
>>> # The Conditional only optimizes the output generation
>>> # for GANs and VAEs, but does NOT guarantee the samples
>>> # are only from that condition.
>>> # If you want to guarantee that output contains only
>>> # "target" == 1 samples, use Constraints.
>>>
>>> constraints = Constraints(
>>>     rules=[
>>>         ("target", "==", 1),
>>>     ]
>>> )
>>> X_gen = test_plugin.generate(count,
>>>         cond=np.ones(count),
>>>         constraints=constraints
>>>        )
>>> assert (X_gen["target"] == 1).all()

Returns

self

classmethod fqdn() → str

The Fully-Qualified name of the plugin.

generate(count: Optional[int] = None, constraints: Optional[synthcity.plugins.core.constraints.Constraints] = None, random_state: Optional[int] = None, **kwargs: Any) → synthcity.plugins.core.dataloader.DataLoader

Synthetic data generation method.

Parameters

  • count – optional int. The number of samples to generate. If None, it generated len(reference_dataset) samples.

  • cond – Optional, Union[pd.DataFrame, pd.Series, np.ndarray]. Optional Generation Conditional. The conditional can be used only if the model was trained using a conditional too. If provided, it must have count length. Not all models support conditionals. The conditionals can be used in VAEs or GANs to speed-up the generation under some constraints. For model agnostic solutions, check out the constraints parameter.

  • constraints –

    optional Constraints. Optional constraints to apply on the generated data. If none, the reference schema constraints are applied. The constraints are model agnostic, and will filter the output of the generative model. The constraints are a list of rules. Each rule is a tuple of the form (<feature>, <operation>, <value>).

    Valid Operations:

    • ”<”, “lt” : less than <value>

    • ”<=”, “le”: less or equal with <value>

    • ”>”, “gt” : greater than <value>

    • ”>=”, “ge”: greater or equal with <value>

    • ”==”, “eq”: equal with <value>

    • ”in”: valid for categorical features, and <value> must be array. for example, (“target”, “in”, [0, 1])

    • ”dtype”: <value> can be a data type. For example, (“target”, “dtype”, “int”)

    Usage example:
    >>> from synthcity.plugins.core.constraints import Constraints
>>> constraints = Constraints(
>>>   rules=[
>>>             ("InterestingFeature", "==", 0),
>>>         ]
>>>     )
>>>
>>> syn_data = syn_model.generate(
        count=count,
        constraints=constraints
    ).dataframe()
>>>
>>> assert (syn_data["InterestingFeature"] == 0).all()

  • random_state – optional int. Optional random seed to use.

Returns

<count> synthetic samples

static hyperparameter_space(**kwargs: Any) → List[synthcity.plugins.core.distribution.Distribution]

Returns the hyperparameter space for the derived plugin.

static load(buff: bytes) → Any
static load_dict(representation: dict) → Any
static name() → str

The name of the plugin.

plot(plt: Any, X: synthcity.plugins.core.dataloader.DataLoader, count: Optional[int] = None, plots: list = ['marginal', 'associations', 'tsne'], **kwargs: Any) → Any

Plot the real-synthetic distributions.

Parameters

  • plt – output

  • X – DataLoader. The reference dataset.

Returns

self

classmethod sample_hyperparameters(*args: Any, **kwargs: Any) → Dict[str, Any]

Sample value from the hyperparameter space for the current plugin.

classmethod sample_hyperparameters_optuna(trial: Any, *args: Any, **kwargs: Any) → Dict[str, Any]
save() → bytes
save_dict() → dict
save_to_file(path: pathlib.Path) → bytes
schema() → synthcity.plugins.core.schema.Schema

The reference schema

schema_includes(other: Union[synthcity.plugins.core.dataloader.DataLoader, pandas.core.frame.DataFrame]) → bool

Helper method to test if the reference schema includes a Dataset

Parameters

other – DataLoader. The dataset to test

Returns

bool, if the schema includes the dataset or not.

training_schema() → synthcity.plugins.core.schema.Schema

The internal schema

static type() → str

The type of the plugin.

static version() → str

API version

class network_edge(feature, parents)

Bases: tuple

count(value, /)

Return number of occurrences of value.

feature

Alias for field number 0

index(value, start=0, stop=9223372036854775807, /)

Return first index of value.

Raises ValueError if the value is not present.

parents

Alias for field number 1

plugin

alias of synthcity.plugins.privacy.plugin_privbayes.PrivBayesPlugin

usefulness_minus_target(k: int, num_attributes: int, num_tuples: int, target_usefulness: int = 5, epsilon: float = 0.1) → int

Usefulness function in PrivBayes.

Parameters

  • k (int) – Max number of degree in Bayesian networks construction

  • num_attributes (int) – Number of attributes in dataset.

  • num_tuples (int) – Number of tuples in dataset.

  • target_usefulness (int or float) –

  • epsilon (float) – Parameter of differential privacy.