graphtoolbox.interpretability.diagnostics

Reusable edge-attribution and ALE diagnostics.

class graphtoolbox.interpretability.diagnostics.RegressionExplanationWrapper(model)

Bases: Module

Expose a GraphToolbox model as a plain node-regression callable.

forward(x, edge_index, **kwargs)

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

graphtoolbox.interpretability.diagnostics.make_edge_explainer(model, epochs: int = 100)

Build a PyG GNNExplainer configured for node regression.

graphtoolbox.interpretability.diagnostics.compute_edge_masks(dataset, model, epochs: int = 100, limit: int | None = None, output_path: str | None = None) → ndarray

Run an edge explainer on graph snapshots and optionally cache its masks.

graphtoolbox.interpretability.diagnostics.topk_indices(values, count: int) → set[int]

Indices of the count largest values.

graphtoolbox.interpretability.diagnostics.explainer_reproducibility(dataset, model, day: int = 0, runs: int = 2, epochs: int = 100) → dict[str, float]

Measure agreement between independent explanations of one snapshot.

graphtoolbox.interpretability.diagnostics.dump_attention_batches(model, dataset, output_dir: str, batch_size: int = 32, require_full_batch: bool = True) → int

Save attention dictionaries for successive PyG mini-batches.

graphtoolbox.interpretability.diagnostics.load_attention_batches(attention_dir: str)

Load cached attention as a [snapshots, edges] tensor.

Heads and message-passing layers are averaged. The returned edge index is checked for consistency across cached batches.

graphtoolbox.interpretability.diagnostics.aggregate_ale_importance(table, group_col: str = 'group', value_col: str = 'importance', reduction: str = 'mean')

Aggregate feature-level ALE importance into normalized group shares.

The default is a mean so groups with many lag columns are not rewarded merely for containing more features. Pass reduction='sum' when total feature importance, rather than typical within-group importance, is desired.

graphtoolbox.interpretability.diagnostics.plot_ale_group_importance(table, output_path: str | None = None, group_col: str = 'group', value_col: str = 'importance', color: str = '#7B2CBF', figsize=(3.3, 1.8))

Render normalized ALE importance by feature group.