PCNOTEARSAlgorithm<T>
Models & Types · AiDotNet.CausalDiscovery.Hybrid
PC-NOTEARS — Hybrid of PC skeleton discovery with NOTEARS continuous optimization.
Inheritance: CausalDiscoveryBase<T> → HybridBase<T> → PCNOTEARSAlgorithm<T>
For Beginners
This hybrid first uses statistical tests to quickly figure out which variable pairs MIGHT be connected, then uses optimization to find the exact edge weights and directions — getting both speed and accuracy.
How It Works
PC-NOTEARS combines the constraint-based PC algorithm’s efficient skeleton discovery with NOTEARS’ continuous optimization for edge weight estimation and orientation. Phase 1 runs PC-style CI tests to identify which variable pairs are connected. Phase 2 runs a NOTEARS-like optimization restricted to the PC skeleton, estimating edge weights while enforcing the acyclicity constraint h(W) = 0.
Algorithm:
- PC skeleton phase: Start with complete graph, remove edges via CI tests
at increasing conditioning set sizes
- NOTEARS phase: Initialize W from PC skeleton (OLS weights for connected pairs,
zero for removed pairs). Run gradient descent on L(W) + lambda * ||W||_1 subject to h(W) = tr(e^{W*W}) - d = 0, but only update entries in the skeleton
- Threshold small weights and return the DAG
Constructors
| Constructor | Summary |
|---|---|
PCNOTEARSAlgorithm(CausalDiscoveryOptions options) |
Initializes PC-NOTEARS with optional configuration. |
Properties
| Property | Summary |
|---|---|
bool SupportsNonlinear { get; } |
|
string Name { get; } |