gsnn.gsnn.proc.construct

construct.py - Build pruned constraint networks for the GSNN model

This helper stitches together several edge lists into a directed heterogeneous graph, removes nodes that do not lie on any path from an input to an output, and returns the result as a PyTorch-Geometric HeteroData object ready for use in GSNN.

Accepted edge tables (all pd.DataFrame with columns src, dst)

  1. input_edges - critical inputs → function nodes

  2. mediator_edges - (optional) mediator inputs → function nodes. Mediators are only retained if they target a function node that survives the pruning step.

  3. function_edges - function → function (latent logic)

  4. output_edges - function → output

ASCII toy graph

``` input_A mediator_M

│ │ ▼ │

func_X ◀────────┘

│ ▼

out_Y

``` mediator_M is kept only if func_X remains on a viable input_A → … → out_Y path.

Quick usage example

```python from gsnn.proc.construct import GSNNNetworkConstructor

builder = GSNNNetworkConstructor(depth=10, verbose=True) data = builder.build(input_edges, function_edges,

output_edges, mediator_edges)

print(data.graph_summary) ```

Returned attributes

  • data.node_names_dict - mapping of node types to names

  • data.edge_index_dict - edge indices (PyTorch tensors)

  • data.graph_summary - pruning statistics & graph metrics

  • data.build_args - parameters used to build the graph

Functions

subset_graph(G, depth, roots, leafs[, ...])

Subset a graph to include only nodes that lie on paths from roots to leaves.

Classes

GSNNNetworkConstructor([depth, verbose])

class gsnn.gsnn.proc.construct.GSNNNetworkConstructor(depth=10, verbose=True)[source]

Bases: object

build(input_edges, function_edges, output_edges, mediator_edges=None, input_names=None, output_names=None, function_names=None, mediator_names=None)[source]

Construct and filter constraint network for GSNN model.

Behavior of fixed node name arguments: - input_names, mediator_names, function_names, output_names specify nodes that MUST

be included in the final node lists, even if they have no incident edges.

  • Additional nodes of each type that are inferred from the provided edge tables are also included. In other words, the final node sets are supersets of the fixed names.

  • Inputs and mediators are combined into a single ‘input’ type in the order input_names + mediator_names, followed by any additional discovered inputs.

Parameters:

  • input_edges – DataFrame with columns ‘src’, ‘dst’ for input→function edges

  • function_edges – DataFrame with columns ‘src’, ‘dst’ for function→function edges

  • output_edges – DataFrame with columns ‘src’, ‘dst’ for function→output edges

  • mediator_edges – Optional DataFrame with columns ‘src’, ‘dst’ for mediator→function edges (only retained if they target a function node retained in the pruned graph)

  • input_names – Optional list of input node names to force-include

  • mediator_names – Optional list of mediator node names to force-include (combined with inputs)

  • function_names – Optional list of function node names to force-include

  • output_names – Optional list of output node names to force-include

Returns:

HeteroData object with filtered network structure and metadata