DYNOTEARSAlgorithm<T>
Models & Types · AiDotNet.CausalDiscovery.TimeSeries
DYNOTEARS — Dynamic NOTEARS for time series structure learning.
Inheritance: CausalDiscoveryBase<T> → TimeSeriesCausalBase<T> → DYNOTEARSAlgorithm<T>
For Beginners
DYNOTEARS is like NOTEARS but for time series. It can learn both “X and Y affect each other at the same time” and “yesterday’s X affects today’s Y” type relationships simultaneously, using the same elegant continuous optimization approach. The key insight is that only the contemporaneous matrix W needs to be acyclic — lagged effects can’t create instantaneous cycles.
How It Works
DYNOTEARS extends the NOTEARS continuous optimization framework to time series data. It jointly learns both contemporaneous (W) and lagged (A₁, …, Aₖ) adjacency matrices using an augmented Lagrangian with the acyclicity constraint only on the contemporaneous matrix W.
Model: X(t) = W^T X(t) + Σ_k A_k^T X(t-k) + e(t) Objective: min_{W,A} ½n⁻¹ ||X_t - X_t W - Z A||²_F + λ₁(||W||₁ + ||A||₁) Constraint: h(W) = tr(e^(W∘W)) - d = 0 (acyclicity only on contemporaneous W)
Reference: Pamfil et al. (2020), “DYNOTEARS: Structure Learning from Time-Series Data”, AISTATS.
Constructors
| Constructor | Summary |
|---|---|
DYNOTEARSAlgorithm(CausalDiscoveryOptions options) |
Properties
| Property | Summary |
|---|---|
bool SupportsNonlinear { get; } |
|
string Name { get; } |