Edit on GitHub

starplast.network_prediction

Family-held-out diffusion over explicitly sourced network layers.

This method is transductive: all graph nodes and fixed edges are visible, but test labels are never seeds or tuning targets. Layer weights and temperature use two different inner validation sets. Unsupported nodes explicitly abstain.

def bundled_network(nodes):

Load the installed graph and declared domain, expression and structure inputs.

Only bundled graphs have these known builders. Custom graphs must use run_network with their own explicit source declarations. Row identity is still verified against the analysis table when the model runs.

def run_network( nodes, spec, graph, layers, edge_sources, log=<built-in function print>):

Evaluate weighted propagation with explicit edge provenance and aligned IDs.

graph must contain ordered gene_ids and the usual layer__a/b/w arrays. edge_sources maps every layer to the columns used to build it; registered target-derived layers are excluded even if omitted from this map. User-supplied edge provenance must describe all label-derived inputs.