dotime.RegimeSwitchingTemporalSCM

class dotime.RegimeSwitchingTemporalSCM(dags, mechanisms, noise, transition_matrix, device=device(type='cpu'), dtype=torch.float32)[source]

Bases: object

Regime-switching temporal SCM.

At each time step, a discrete regime variable d_t ∈ {0,…,R-1} determines which causal structure and mechanisms are active:

X_t^(i) = f_i^(d_t)(Pa_{G^(d_t)}(X_t^(i))) + ε_t^(i)

Regime transitions follow a Markov chain: d_t ~ Categorical(P[d_{t-1}, :])

Parameters:
__init__(dags, mechanisms, noise, transition_matrix, device=device(type='cpu'), dtype=torch.float32)[source]
Parameters:

  • dags (List[TemporalDAG]) – List of temporal DAGs, one per regime.

  • mechanisms (List[Dict[str, TemporalMechanism]]) – List of mechanism dictionaries, one per regime.

  • noise (Dict[str, DistributionSampler]) – Noise distributions (shared across regimes).

  • transition_matrix (np.ndarray) – Regime transition matrix P of shape (R, R).

  • device (torch.device) – Device for computation.

  • dtype (torch.dtype) – Data type.

Methods

__init__(dags, mechanisms, noise, ...[, ...])

sample_interventional(T, intervention[, ...])

Sample interventional data with regime switching.

sample_observational(T[, burn_in, ...])

Sample observational data with regime switching.
sample_observational(T, burn_in=50, generator=None, return_regimes=False)[source]

Sample observational data with regime switching.

Parameters:

  • return_regimes (bool) – If True, return (X, regimes) tuple. Otherwise just X.

  • T (int)

  • burn_in (int)

  • generator (Generator | None)

Return type:

Tensor

Returns:

torch.Tensor or Tuple[torch.Tensor, np.ndarray] – Time series (T, N), or (time series, regimes) if return_regimes=True.

sample_interventional(T, intervention, burn_in=50, generator=None, return_regimes=False)[source]

Sample interventional data with regime switching.

Parameters:
Return type:

Tensor

Returns:

torch.Tensor or Tuple[torch.Tensor, np.ndarray] – Time series (T, N), or (time series, regimes) if return_regimes=True.