NTSNOTEARSAlgorithm<T>
Models & Types · AiDotNet.CausalDiscovery.TimeSeries
NTS-NOTEARS — Nonstationary Time Series NOTEARS.
Inheritance: CausalDiscoveryBase<T> → TimeSeriesCausalBase<T> → NTSNOTEARSAlgorithm<T>
For Beginners
Regular time series methods assume the causal relationships stay the same forever. NTS-NOTEARS can detect when relationships change — for example, a market regime shift where the causes of stock prices change.
How It Works
NTS-NOTEARS extends DYNOTEARS to handle nonstationary time series where the causal structure may change over time. It partitions the data into segments using a variance-based change-point detector, learns a separate NOTEARS-style DAG for each segment, and produces a summary graph by taking the maximum absolute edge weight across segments, preserving regime-specific causal effects.
Algorithm:
- Partition time series into K segments using variance-based change-point detection
- For each segment, learn a contemporaneous DAG via NOTEARS with DYNOTEARS-style
lagged terms and the augmented Lagrangian acyclicity constraint
- Aggregate segment-level DAGs into a summary graph: edge weight = max |w_k| over
segments where that edge appears, preserving regime-specific causal effects
Reference: Sun et al. (2021), “NTS-NOTEARS: Learning Nonparametric DBN Structure from Nonstationary Time Series”.
Constructors
| Constructor | Summary |
|---|---|
NTSNOTEARSAlgorithm(CausalDiscoveryOptions options) |
Properties
| Property | Summary |
|---|---|
bool SupportsNonlinear { get; } |
|
string Name { get; } |