Mechanism abstract base class — Mechanism • kagu

A mechanism owns the full lifecycle of a node's conditional model, decoupled from any particular inference backend. Subclasses implement:

$fit(node, parents, data, ...) - fit the local model, returning an opaque fit object.
$predict_mean(node, parents, parent_values, fit) - the conditional mean E[node | parents] for each posterior draw, as a [n_chains, n_draws] matrix (Gaussian processes use a single chain, so n_chains = 1).
$log_marglik(node, parents, data) - the log marginal likelihood log P(node | parents), used by structure discovery.
$posterior_shape(fit) - c(n_chains, n_draws) for the fit.
$node_terms(node, parents, data, fit) - summary rows for model$summary().

Methods

Fit the node's local model.

Mechanism$fit(node, parents, data, ...)

An opaque fit object.

Conditional mean for each posterior draw.

Mechanism$predict_mean(node, parents, parent_values, fit)

node: Character scalar - the node name.
parents: Character vector of parent node names.
parent_values: Named list of [n_chains, n_draws] matrices, one per parent, giving the parent values for each posterior draw.
fit: A fit object from $fit().

A [n_chains, n_draws] numeric matrix.

Log marginal likelihood log P(node | parents).

Mechanism$log_marglik(node, parents, data)

A single numeric.

Posterior shape of a fit.

Mechanism$posterior_shape(fit)

Named integer vector c(n_chains, n_draws).

Summary rows for this node (used by model$summary()).

Mechanism$node_terms(node, parents, data, fit)

A tibble with columns node, term, mean, sd, hdi_lower, hdi_upper.

The objects of this class are cloneable with this method.

Mechanism$clone(deep = FALSE)