dotime.TemporalSCM

class dotime.TemporalSCM(dag, mechanisms, noise, device=device(type='cpu'), dtype=torch.float32)

Bases: object

Temporal Structural Causal Model with time-stepped forward simulation.

Extends Do-PFN’s SCM to support temporal dependencies with lags.

Parameters:

• dag (TemporalDAG)

• mechanisms (dict[str, TemporalMechanism])

• noise (dict[str, DistributionSampler])

• device (device)

• dtype (dtype)

__init__(dag, mechanisms, noise, device=device(type='cpu'), dtype=torch.float32)

Parameters:

• dag (TemporalDAG) – Temporal DAG with instantaneous and lagged edges.

• mechanisms (Dict[str, TemporalMechanism]) – Mechanisms for each variable.

• noise (Dict[str, DistributionSampler]) – Noise distributions for each variable.

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

• dtype (torch.dtype) – Data type for computation.

Methods

__init__(dag, mechanisms, noise[, device, dtype])
sample_interventional(T, intervention[, ...])	Sample interventional data from the temporal SCM.
sample_observational(T[, burn_in, generator])	Sample observational data from the temporal SCM.

sample_observational(T, burn_in=50, generator=None)

Sample observational data from the temporal SCM.

Parameters:

• T (int) – Length of time series to generate (after burn-in).

• burn_in (int) – Number of burn-in steps to discard.

• generator (torch.Generator, optional) – RNG for reproducibility.

Return type:

Tensor

Returns:

torch.Tensor – Time series data of shape (T, N) where N is number of variables.

sample_interventional(T, intervention, burn_in=50, generator=None)

Sample interventional data from the temporal SCM.

Parameters:

• T (int) – Length of time series to generate (after burn-in).

• intervention (InterventionSpec) – Intervention specification.

• burn_in (int) – Number of burn-in steps to discard.

• generator (torch.Generator, optional) – RNG for reproducibility.

Return type:

Tensor

Returns:

torch.Tensor – Time series data of shape (T, N) under intervention.