Changelog¶

All notable changes to graphdml are listed here. Versions follow semantic versioning; before 1.0, minor versions may change behavior.

Unreleased ¶

Documentation site (https://samrdreyfuss.github.io/graphdml/): guides, methodology, validation, example datasets and an API reference generated from the docstrings.
graphdml.selftest() and make_signal_check(): a built-in check, on synthetic data with planted effects and a null twin that has none, that the estimator finds real signals and does not invent false ones. Reports bias, coverage, power and false-alarm rates against fixed criteria, next to a network-blind baseline.
ATE (total effect): the average effect of treating every node vs none, reported first in summary() and as a "total" row in summary_frame(), cluster_summary_frame() and conf_int(), with model.ate_ and model.total_weights_. Pass total=False for the direct and peer rows only.
Practitioner-facing labels in summary(): ATE (total), direct (ADE), peer (APE), with a plain-language unit line for each and a note on z-statistics.
GraphData.with_values(..., feature_names=...); covariates may change their number of columns.
Installation guide, badges and project links in the README and on PyPI.
README datasets table now says which datasets need outside data (LastFM downloads on first use, the insurance experiment must be downloaded from openICPSR), and lists make_toy_graph.

Repository polish: code of conduct, security policy, issue forms and a pull-request template; py.typed marker so type checkers use graphdml's annotations; CI pulls the Neo4j test image through a mirror, which fixes random failures from Docker Hub rate limits.
load_insurance_experiment explains where to get the data when the files are missing, and fetch_lastfm_asia explains what to do when the download fails.

summary_frame() includes a "total" row by default when a peer exposure is fitted.
Removed the unused gnn and viz optional extras.

The placebo test in the test suite is seeded, so it no longer fails occasionally.

First release.

GraphDML estimator for direct and peer effects on networks: focal sets from dependency sets, cross-fitting with buffered nuisance training, IV-type and partialling-out scores, sandwich and cluster-robust standard errors, sub-population estimation, and mode="original" for the original GDML procedure.
Nuisance models from any scikit-learn estimator, with neighborhood features.
Diagnostics and falsification checks: placebo_test, negative_control_test, two_hop_test.
Example datasets with known truth, a semi-synthetic LastFM Asia dataset, and a loader for the Cai, de Janvry & Sadoulet (2015) randomized network experiment.
Neo4j integration: load graphs, write results back, GDS FastRP features, demo.
Validation suite and benchmarks.