synthcity.plugins.core.models.mbi.inference module

class FactoredInference(domain, structural_zeros={}, metric='L2', log=False, iters=1000, warm_start=False, elim_order=None)

Bases: object

dual_averaging(measurements, total=None, lipschitz=None, callback=None)
Use the regularized dual averaging algorithm to estimate the GraphicalModel

See https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/xiao10JMLR.pdf

Parameters

  • measurements – a list of (Q, y, noise, proj) tuples, where Q is the measurement matrix (a numpy array or scipy sparse matrix or LinearOperator) y is the noisy answers to the measurement queries noise is the standard deviation of the noise added to y proj defines the marginal used for this measurement set (a subset of attributes)

  • total – The total number of records (if known)

  • lipschitz – the Lipchitz constant of grad L(mu) - automatically calculated for metric=L2 - doesn’t exist for metric=L1 - must be supplied for custom callable metrics

  • callback – a function to be called after each iteration of optimization

estimate(measurements, total=None, engine='MD', callback=None, options={})

Estimate a GraphicalModel from the given measurements

Parameters

  • measurements – a list of (Q, y, noise, proj) tuples, where Q is the measurement matrix (a numpy array or scipy sparse matrix or LinearOperator) y is the noisy answers to the measurement queries noise is the standard deviation of the noise added to y proj defines the marginal used for this measurement set (a subset of attributes)

  • total – The total number of records (if known)

  • engine – the optimization algorithm to use, options include: MD - Mirror Descent with armijo line search RDA - Regularized Dual Averaging IG - Interior Gradient

  • callback – a function to be called after each iteration of optimization

  • options – solver specific options passed as a dictionary { param_name : param_value }

Return model

A GraphicalModel that best matches the measurements taken

fix_measurements(measurements)
infer(measurements, total=None, engine='MD', callback=None, options={})
interior_gradient(measurements, total, lipschitz=None, c=1, sigma=1, callback=None)
Use the interior gradient algorithm to estimate the GraphicalModel

See https://epubs.siam.org/doi/pdf/10.1137/S1052623403427823 for more information

Parameters

  • measurements – a list of (Q, y, noise, proj) tuples, where Q is the measurement matrix (a numpy array or scipy sparse matrix or LinearOperator) y is the noisy answers to the measurement queries noise is the standard deviation of the noise added to y proj defines the marginal used for this measurement set (a subset of attributes)

  • total – The total number of records (if known)

  • lipschitz – the Lipchitz constant of grad L(mu) - automatically calculated for metric=L2 - doesn’t exist for metric=L1 - must be supplied for custom callable metrics

  • sigma (c,) – parameters of the algorithm

  • callback – a function to be called after each iteration of optimization

mirror_descent(measurements, total=None, stepsize=None, callback=None)
Use the mirror descent algorithm to estimate the GraphicalModel

See https://web.iem.technion.ac.il/images/user-files/becka/papers/3.pdf

Parameters

  • measurements – a list of (Q, y, noise, proj) tuples, where Q is the measurement matrix (a numpy array or scipy sparse matrix or LinearOperator) y is the noisy answers to the measurement queries noise is the standard deviation of the noise added to y proj defines the marginal used for this measurement set (a subset of attributes)

  • stepsize – The step size function for the optimization (None or scalar or function) if None, will perform line search at each iteration (requires smooth objective) if scalar, will use constant step size if function, will be called with the iteration number

  • total – The total number of records (if known)

  • callback – a function to be called after each iteration of optimization