Backends

compile_graph() compiles the same source / target edgelist into one of the supported sparse PyTorch models defined by the backend. The backend controls how the model is executed.

Feedforward

Compiles an acyclic, layerable graph into successive masked layer blocks: a sparse multilayer perceptron whose connectivity comes from the edgelist.

Skip edges are expanded internally with pseudo nodes so every compiled hop is adjacent layer-to-layer. Pseudo nodes never appear in artifact.interpretation_sites or in node attribution.

Omit steps. Feature and node attribution both work. Node sites are layer_1, layer_2, and so on, starting after the input layer, so inputs are not in result.sites (use result.table).

Use this when the graph is hierarchical. It will not compile cyclic graphs or directed graphs that cannot be layered.

State-update

Keeps the original topology, including cycles, and applies a shared masked update for a fixed number of steps (default 3). steps is the unrolling count per forward pass, not a training epoch count and not a sequence length in the data. Passing steps with backend="feedforward" raises.

Each node is a scalar. Vector or matrix nodes are currently not supported. After every step, input-node values are re-injected. Every node that can influence an output must be reachable from at least one input. Read named edge weights with edge_weights(model).

Feature and node attribution both work. Node sites are step_1, step_2, and so on, and include all visible nodes, including inputs. result.table and result.sites store signed Captum scores. result.summary(aggregation=...) collapses those repeated columns (peak, mean_abs, or last). "peak" (default) keeps the signed score from the step with largest magnitude; "last" is also signed; "mean_abs" averages magnitudes.

Use this for cycles or feedback. It is not an LSTM or sequence model, does not learn different weights per unrolled step, and does not run a converge-until loop inside compile_graph().