Changelog¶

All notable changes to this project will be documented in this file.

This project follows semantic versioning.

[0.3.0] - in preparation¶

API redesign. Same job as 0.2.0: edgelist → sparse PyTorch module → optional Captum attribution. Call sites and return types change.

0.2.0	0.3.0
`backend="recurrent"` or `"graphnn"`	`backend="state_update"`
`model.recurrent` / `model.message_passing`	`model.state_linear`
`interpret_model(..., target=)`	`interpret_model(..., on=)`
Captum output index also named `target=`	`attribute_kwargs={"target": ...}`
`interpret_model` returned a DataFrame or a dict of tables	always `InterpretationResult`
`level=` / `nodes=` / `site_aggregation=` on `interpret_model`	`result.summary(nodes=..., aggregation=...)`
`site_aggregation="max_abs"`	`aggregation="peak"` (no alias; keeps the signed score from the step with largest \|score\|)
feature table / node tables	`result.table` / `result.sites`
`graph_topology()`	`artifact.backend`, `input_nodes`, `output_nodes`, `hidden_nodes`, `interpretation_sites` (`feature_names` is an alias of `input_nodes`)
`model.*.weight`	`edge_weights(model)` (parameters are `*.raw_weight`; checkpoints from 0.2.0 will not load as-is)
`quiet=`	removed (`compile_graph` notes go to the `edge2torch` logger)
`compile_graph(..., backend="feedforward", steps=...)` ignored `steps`	now raises; omit `steps` on feedforward (default remains 3 on `state_update`)

Catch public failures with from edge2torch import Edge2TorchError.

Breaking: the state-update summary aggregation "max_abs" is renamed "peak". The default still keeps the signed score from the step with largest magnitude. There is no "max_abs" alias. "last" is also signed; only "mean_abs" drops the sign.

Extra DataFrame columns and AnnData variables are dropped, not errors.
Omitting method uses IntegratedGradients for on="features" and LayerConductance for on="nodes".
Attribution tables store signed Captum scores. Rankings or plots that want magnitude should call .abs() themselves.

Node interpretation on all backends. You can now attribute predictions to named hidden nodes on recurrent and graphnn models, not only on feedforward.
Finer control over node attribution output via new interpret_model() options:
level="summary" — one table per sample (default)
level="sites" — separate tables per interpretation site (layer_* on feedforward, step_* on recurrent and graphnn)
nodes — include hidden nodes only, or also outputs ("non_input"), or all visible nodes ("all")
site_aggregation — on recurrent and graphnn, choose how step-wise scores are combined in the summary ("max_abs", "mean_abs", or "last")
New example notebooks: recurrent and graphnn end-to-end workflows (compile, train, interpret on cyclic graphs).
New docs page: Scope and limitations — what each backend supports cleanly, with extra PyTorch work, or not at all.

Breaking: interpret_model(..., target="nodes") now returns a summary pandas.DataFrame by default. For the previous per-site dictionary of tables, pass level="sites".

GraphNN example notebook: corrected graph topology so the signal path reaches the readout node as intended.

Initial release of edge2torch.
Added compile_graph() for compiling named edge lists into PyTorch models.
Added support for the feedforward, recurrent, and graphnn backends.
Added feature alignment with align_features_to_input_nodes().
Added model customization with customize_model().
Added Captum-based interpretation with interpret_model(), including feature-level attribution and feedforward node-level attribution.
Added documentation, examples, and tests.