LPCMCIAlgorithm<T>
Models & Types · AiDotNet.CausalDiscovery.TimeSeries
LPCMCI — Latent PCMCI for time series with hidden confounders.
Inheritance: CausalDiscoveryBase<T> → TimeSeriesCausalBase<T> → LPCMCIAlgorithm<T>
For Beginners
LPCMCI is the most advanced version of PCMCI. It works even when there are hidden variables affecting the ones you can measure. The trade-off is that some edges may be uncertain in direction (shown as symmetric, bidirected links).
How It Works
LPCMCI extends PCMCI to handle latent confounders by combining ideas from FCI (ancestral graph representation) with PCMCI’s condition selection and MCI testing. It outputs a time series PAG (partial ancestral graph) instead of a DAG.
Algorithm:
- Run PCMCI condition selection to find preliminary lagged parents for each variable
- Run MCI tests: confirm each candidate link X_i(t-τ) → X_j(t) conditioning on the
parents of BOTH endpoints, removing links that become independent
- Apply orientation rules that account for possible latent confounders:
temporal (past→present) orientation, plus symmetric bidirected marks where both directions survive (ambiguous / latent-confounded)
- Compute summary graph: collapse lags, keep max absolute strength per pair
Reference: Gerhardus and Runge (2020), “High-recall causal discovery for autocorrelated time series with latent confounders”, NeurIPS.
Constructors
| Constructor | Summary |
|---|---|
LPCMCIAlgorithm(CausalDiscoveryOptions options) |
Properties
| Property | Summary |
|---|---|
bool SupportsLatentConfounders { get; } |
|
bool SupportsNonlinear { get; } |
|
string Name { get; } |