NOTEARSLowRank<T>
Models & Types · AiDotNet.CausalDiscovery.ContinuousOptimization
NOTEARS Low-Rank — DAG learning with low-rank parameterization for scalability.
Inheritance: CausalDiscoveryBase<T> → ContinuousOptimizationBase<T> → NOTEARSLowRank<T>
For Beginners
When you have many variables (say hundreds), the standard NOTEARS weight matrix becomes very large. The low-rank trick says “most causal graphs are relatively simple” and represents the matrix using fewer numbers, making the optimization much faster.
How It Works
Parameterizes the weighted adjacency matrix W as a product of low-rank factors W = A * B^T, where A and B are d x r matrices with r much less than d. This reduces the number of parameters from O(d^2) to O(dr) and enables scalability to graphs with many variables. The NOTEARS acyclicity constraint h(W) = tr(e^(W * W)) - d is applied to the reconstructed W.
Algorithm:
- Choose rank r = min(d, 10)
- Initialize A, B as small random matrices of size d x r
- Reconstruct W = A * B^T (with diagonal zeroed)
- Compute L2 loss and NOTEARS acyclicity constraint on W
- Compute gradients via chain rule: dL/dA = dL/dW * B, dL/dB = dL/dW^T * A
- Update A, B via gradient descent with augmented Lagrangian for acyclicity
- Threshold the final W = A * B^T
Reference: Fang et al. (2023), “On Low-Rank Directed Acyclic Graphs and Causal Structure Learning”, IEEE Transactions on Signal Processing.
Constructors
| Constructor | Summary |
|---|---|
NOTEARSLowRank(CausalDiscoveryOptions options) |
Initializes NOTEARS Low-Rank with optional configuration. |
Properties
| Property | Summary |
|---|---|
bool SupportsNonlinear { get; } |
|
string Name { get; } |