BayesianNetworkSynthGenerator<T>
Models & Types · AiDotNet.NeuralNetworks.SyntheticData
PrivBayes synthesizer (Zhang et al., SIGMOD 2014 / TODS 2017): learns a k-degree Bayesian network under differential privacy, releases noisy conditional distributions, and generates synthetic data by ancestral sampling.
Inheritance: SyntheticTabularGeneratorBase<T> → BayesianNetworkSynthGenerator<T>
For Beginners
Think of this as building a probabilistic “family tree” of your features, then adding carefully calibrated random noise so that no single person’s record can be recovered from the released model.
How It Works
With EnableDifferentialPrivacy on (the default) the two
phases follow the paper:
- GreedyBayes (Algorithm 1 / 4): the first attribute is chosen uniformly at random. Each later
attribute X is attached together with a parent set of exactly min(k, |V|) already-attached
attributes, chosen by the exponential mechanism over the mutual information I(X; Pi). The sensitivity
is Lemma 1’s binary form when X or Pi has a binary domain and the general form otherwise.
- NoisyConditionals (Algorithm 2): for i = k+1..d the full joint distribution of
(X_i, Pi_i), zero cells included, receives Laplace noise of scale 2(d-k)/(n eps2); negatives are
clipped and the table renormalized. The conditionals of the first k attributes are derived from the
noisy joint of (X_{k+1}, Pi_{k+1}), so exactly d-k marginals spend the phase-2 budget.
The degree k is the largest value up to MaxParents that
keeps the network theta-useful (Section 4.4): n eps2 / (2 (d-k) |dom(X, Pi)|) >= theta.
Constructors
| Constructor | Summary |
|---|---|
BayesianNetworkSynthGenerator(BayesianNetworkSynthOptions<T> options) |
Initializes a new instance of the BayesianNetworkSynthGenerator class. |