AIMOptions<T>

Options & Configuration · AiDotNet.Models.Options

Configuration options for AIM (Adaptive Iterative Mechanism), a marginal-based differentially private synthetic data generation method.

Inheritance: ModelOptions → NeuralNetworkOptions → ModelHyperparameterOptions → FinancialNeuralNetworkOptions → RiskModelOptions<T> → AIMOptions<T>

For Beginners

AIM is a non-neural approach to private data synthesis:

  1. Pick important “statistics” about the data (marginals = histograms of 1-3 columns)
  2. Measure them with added noise (for privacy)
  3. Create fake data that matches those noisy statistics

Unlike GAN/VAE models, AIM uses mathematical optimization (not deep learning), which often works better for smaller datasets and provides formal privacy guarantees.

Example:

How It Works

AIM (McKenna et al. 2022) generates synthetic data through adaptive, iterative marginal measurements under zero-concentrated differential privacy:

  • Marginal selection: the exponential mechanism picks the worst-approximated workload marginal
  • Noisy measurement: the Gaussian mechanism measures it; the noise is annealed as the model improves
  • Model estimation: Private-PGM fits a graphical model to every noisy measurement

Reference: “AIM: An Adaptive and Iterative Mechanism for Differentially Private Synthetic Data” (McKenna et al., 2022)

Constructors

Constructor Summary
AIMOptions()

Properties

Property Summary
Vector<T> DomainLowerBounds { get; set; } Gets or sets the public lower bound of each column’s domain, used for the equal-width bins.
Vector<T> DomainUpperBounds { get; set; } Gets or sets the public upper bound of each column’s domain, used for the equal-width bins.
double Delta { get; set; } Gets or sets delta of the (epsilon, delta) guarantee.
double Epsilon { get; set; } Gets or sets the total privacy budget (epsilon).
double MaxModelSizeMb { get; set; } Gets or sets the maximum size of the graphical model’s junction tree, in megabytes.
double MeasurementBudgetFraction { get; set; } Gets or sets alpha, the fraction of each round’s zCDP budget spent on measurement (the rest pays for selection).
int MaxCells { get; set; } Gets or sets the largest number of cells a workload marginal may have.
int MaxMarginalOrder { get; set; } Gets or sets the order of the workload marginals (every combination of this many columns).
int NumBins { get; set; } Gets or sets the number of equal-width bins for discretizing numeric columns.
int NumIterations { get; set; } Gets or sets the number of Private-PGM mirror-descent iterations per model estimate.
int? Rounds { get; set; } Gets or sets the number of rounds T used to set the initial noise scales.