Skip to content

Structure Learning Hyperparameters

This table below describes the hyperparameters that are supported by one or more CausalIQ structure learning algorithms.

Hyperparameter Type Description
alpha constraint p-value threshold used in Conditional Independence tests
ci_test constraint Conditional Independence tests used in constraint-based learning
iss constraint Imaginary Sample Size used to weight priors in Bayesian scores
max_elapsed general Maximum elapsed time allowed in seconds
max_iterations score Limit on the number of iterations in score-based learning
no_increase score Number of iterations allowed where the score does not increase
penalty_weight score Weighting assigned to the penalty component of scores such as AIC and BIC
score score Objective score used in score-based learning
tabulist_len score Length of the Tabu list

General hyperparameters

max_elapsed

Maximum allowed elapsed execution time in seconds specified as a positive integer.

Score-based hyperparameters

iss

A real value defining the Imaginary Sample Size (ISS) used to weight the prior in Bayesian scores.

max_iterations

A positive integer placing a limit on the number of iterations.

no_increase

A non-negative integer if iterations where the score is allowed not to increase.

penalty_weight

A positive real number weighting the penalty component of the AIC and BIC scores relative to the log. likelihood.

score

Defines the objective score used by score-based algorithms and the score-based phases of hybrid algorithms. Supported values are:

  • aic: Aikike Information Criterion
  • bdeu: Bayesian Equivalent Uniform
  • bge: Bayesian Gaussian Equivalent
  • bic: the Bayesian Information Criterion
  • k2: K2
  • loglik: Log Likelihood

tabulist_len

A positive integer defining the length of the Tabu list which contains the highest-scoring recently visited graphs.

Constraint-based hyperparameters

alpha

A positive real-value below 1.0 which defines the p-value threshold below which the CI test is interpreted as indication conditional independence.

ci_test

Defines the conditional independence (CI) test used by constraint-based algorithms and the constraint-based phases of hybrid algorithms. Supported values are:

  • mi: Mutual Independence test
  • x2: Chi-squared test