CausalDiscoveryOptions

Options & Configuration · AiDotNet.Models.Options

Configuration options for causal structure discovery.

Inheritance: ModelOptions → CausalDiscoveryOptions

For Beginners

You can use these options to configure how the causal discovery algorithm works. If you leave everything as null, sensible defaults will be used. The most important option is Algorithm which determines which method is used.

Example:

How It Works

These options control which causal discovery algorithm is used and how it behaves. All properties are nullable with industry-standard defaults applied internally when null.

Constructors

Constructor Summary
CausalDiscoveryOptions() Initializes a new instance with default values (all options null → each algorithm’s own default applies).
CausalDiscoveryOptions(CausalDiscoveryOptions other) Initializes a new instance by copying every option from another instance.

Properties

Property Summary
CausalDiscoveryAlgorithmType? Algorithm { get; set; } Which causal discovery algorithm to use.
bool? UseForFeatureSelection { get; set; } Whether to also use the discovered causal graph for feature selection in preprocessing.
bool? UseKlWarmUp { get; set; } Whether to use KL weight warm-up schedule.
double? AcyclicityTolerance { get; set; } Convergence tolerance for the acyclicity constraint h(W).
double? ConcavityParameter { get; set; } Concavity parameter (gamma) for MCP/SCAD penalty functions.
double? ConfoundingEvidenceCutoff { get; set; } Fraction (0..1) of a candidate’s direction-evidence that may point the “wrong” way before RCD treats the remaining variables as latently confounded and stops.
double? CorrelationThreshold { get; set; } Correlation threshold for edge inclusion in constraint-based time-series methods.
double? DefaultKlWeight { get; set; } Default KL divergence weight for variational methods (GAE, CausalVAE).
double? DirectionalityAsymmetryThreshold { get; set; } Minimum relative asymmetry (in [0, 1]) between the two cross-map skills of a pair before CCM prunes the weaker reverse direction as a reconstruction artifact.
double? EdgeThreshold { get; set; } Edge weight threshold for pruning.
double? InitScale { get; set; } Initialization scale for weight matrices or low-rank factors.
double? InitialLogVariance { get; set; } Initial log-variance for variational parameters in GAE and similar methods.
double? LearningRate { get; set; } Learning rate for deep learning-based causal discovery methods.
double? MaxKlWeight { get; set; } Maximum KL divergence weight after warm-up.
double? MaxPenalty { get; set; } Maximum penalty parameter (rho_max) for augmented Lagrangian methods.
double? SignificanceLevel { get; set; } Significance level for conditional independence tests.
double? SobolevWeight { get; set; } Sobolev regularization weight for NOTEARS Sobolev.
double? SparsityPenalty { get; set; } L1 sparsity penalty (lambda1).
double? TCdfAttentionLearningRateMultiplier { get; set; } Multiplier applied to TCDF’s attention-logit learning rate after its dimension-squared Jacobian compensation.
int? HiddenUnits { get; set; } Number of hidden units in neural network layers for deep learning methods.
int? InnerIterations { get; set; } Number of inner gradient descent steps per outer augmented Lagrangian iteration.
int? MaxConditioningSetSize { get; set; } Maximum conditioning set size for constraint-based methods.
int? MaxEpochs { get; set; } Maximum number of training epochs for deep learning-based methods.
int? MaxIterations { get; set; } Maximum number of outer iterations for optimization-based methods.
int? MaxLag { get; set; } Maximum lag order for time-series causal discovery methods.
int? MaxParents { get; set; } Maximum number of parents per node.
int? MaxRank { get; set; } Maximum rank for low-rank matrix factorizations.
int? MaxSegments { get; set; } Maximum number of segments for nonstationary time-series methods.
string LossType { get; set; } Loss type for continuous optimization methods.
string[] FeatureNames { get; set; } Variable/feature names to label the graph nodes.