[vc1492a]

PyNomaly 1.0.0 makes LocalOutlierProbability a proper scikit-learn estimator. This PR carries the whole change: the original refactor plus #92 and #93 from @SaritNike, which finished the compliance work and added a public API contract test.

What changed for users

• Data moves out of the constructor. LocalOutlierProbability(extent, n_neighbors, use_numba, n_jobs, progress_bar) takes configuration only; data, distance_matrix, neighbor_matrix, and cluster_labels go to fit(X, y=None, ...). Passing them to the constructor by keyword still works with a FutureWarning. Passing data positionally to the constructor no longer works, since the first positional parameter is now extent.
• The class inherits BaseEstimator and OutlierMixin, so get_params/set_params, fit_predict, predict, decision_function, and score_samples follow the sklearn outlier conventions (decision_function = score_samples - offset_, inliers are >= 0). check_estimator(LoOP()) passes with one documented xfail, check_fit1d, because LoOP deliberately accepts 1-d input and reshapes it to (-1, 1). scikit-learn is an optional dependency (pip install PyNomaly[sklearn]); a minimal fallback keeps the package importable without it.
• Fitted state lives on _-suffixed attributes (local_outlier_probabilities_, prob_distances_, distance_matrix_, n_neighbors_, ...). Read-only aliases without the underscore remain for the output attributes. Note that distance_matrix after fit() is the constructor input (usually None); the computed matrix is only on distance_matrix_.
• Exceptions now live in PyNomaly.exceptions and are re-exported from PyNomaly and PyNomaly.loop, so existing imports keep working.
• LoOP is an alias for LocalOutlierProbability, and __version__ is exported from the package.
• fit() can be called repeatedly with different data; state is reset each time.
• scipy sparse matrices are accepted by fit, predict, decision_function, and stream (densified internally; scipy is a soft dependency).
• stream() returns a Python float instead of a NumPy array.
• Python 3.8 support is dropped; python_requires='>=3.9', classifiers for 3.9 through 3.14.
• CI runs on every branch and pull request, and the flake8 syntax/undefined-name check now fails the job.

Migration

Before (0.4.x):

m = LocalOutlierProbability(data, extent=2, n_neighbors=20, cluster_labels=labels).fit()
scores = m.local_outlier_probabilities

After (1.0.0):

m = LoOP(extent=2, n_neighbors=20).fit(data, cluster_labels=labels)
scores = m.local_outlier_probabilities_

Deliberately deferred to the follow-up warnings PR

• Removing the deprecated constructor data keywords (they warn today).
• fit() still warns and returns an unfitted model on invalid input instead of raising ValueError; kept for backward compatibility for now.
• Documenting the non-underscore read aliases as deprecated, and eventually warning on access.
• cluster_labels_u became private (_cluster_labels_u); it needs a changelog note if we consider it part of the public API.

Tests as of the compliance head 43ea6e4: 52 passed, 4 skipped locally; CI green on 3.9 through 3.14.

Thanks to @SaritNike for #92 (the compliance work and check_estimator integration) and #93 (the public API contract test).

Closes #7

[Copilot]

Pull request overview

This PR refactors PyNomaly’s LocalOutlierProbability estimator toward scikit-learn-style usage by moving data-bearing inputs from __init__() to fit(), while preserving backward compatibility via FutureWarnings and updating the public API/docs for the 1.0.0 release.

Changes:

• Shift data parameters (data, distance_matrix, neighbor_matrix, cluster_labels) to fit() and add FutureWarning-based backward compatibility for passing them to __init__().
• Refactor the implementation into mixins/modules (_validation, _distance, _pipeline, _utils) and add LoOP alias + __version__ export.
• Update tests and documentation/examples to the new API and bump version references to 1.0.0.

Reviewed changes

Copilot reviewed 14 out of 14 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
tests/test_loop.py	Updates existing tests to call fit(X) and adds new tests for compatibility/refit/alias behavior.
setup.py	Bumps package version to 1.0.0 and updates download URL.
README.md	Updates badges and usage examples to the new fit()-centric API and LoOP alias.
PyNomaly/loop.py	Implements the new estimator API shape (data args in fit()), deprecation warnings, refit reset behavior, and LoOP alias.
PyNomaly/exceptions.py	Extracts custom exceptions into a dedicated module.
PyNomaly/_validation.py	Adds validation mixin + accepts decorator used by the estimator.
PyNomaly/_utils.py	Extracts progress bar helper into a utility module.
PyNomaly/_pipeline.py	Extracts the LoOP scoring pipeline computations into a mixin.
PyNomaly/_distance.py	Extracts distance/neighbor computation into a mixin (vectorized + Numba paths).
PyNomaly/init.py	Re-exports LoOP and __version__ (plus exceptions) as public API.
docs/user-guide.md	Updates documentation code snippets to pass data to fit().
docs/getting-started.md	Updates quickstart snippets to use LoOP and fit(data).
docs/changelog.md	Adds 1.0.0 changelog entry describing the API change and additions.
changelog.md	Mirrors the 1.0.0 changelog entry at repo root.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[SaritNike]

There are many compatibility issues with scikit-learn. Most of them are easy to fix and scikit learn provides helper functions to check the compatibility first. Instead of listing them all here, i will add a commit that includes this check in the unit tests.

[SaritNike]

Scikit learn requires following a certain inheritance structure (e.g. BaseEstimator) etc.
As it stands the code is currently not compatible with scikit-learn API

Dismissing the June review: its findings were addressed in #92, now merged into this branch as a35b62e. Re-requesting review on the combined 1.0.0 change.