Changelog¶

All notable changes to this project are documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Unreleased ¶

Python 3.14 support, tested in CI on Linux, macOS and Windows.
A CI job that installs the lowest versions pyproject.toml allows for every direct dependency (numpy 1.26, scikit-learn 1.6) on Python 3.10 and runs the whole test suite.
ROADMAP.md, also on the documentation site: the milestones to 1.0.0 with exit criteria, measured performance targets and the decisions to take before the API freeze.

Breaking: points tied at the distance of the k-th neighbour share the remaining votes, instead of the neighbour search picking some of them. Every subspace model, in estimators_ and behind the leave-one-out and k-fold votes, is now a TieSharingKNeighborsClassifier rather than a KNeighborsClassifier. Predictions change wherever such ties occur. On the fourteen benchmark datasets, mean macro-F1 of the default ensemble moves from 0.780 to 0.777, almost all of it on datasets with many repeated values and imbalanced classes: blood-transfusion 0.598 to 0.579, ilpd 0.565 to 0.552, qsar-biodeg 0.809 to 0.801. A shared vote is the average over every way of breaking the tie, which leans towards the majority class. docs/benchmark.md has every dataset.
The greedy selection treats losses within 1e-12 of each other as equal and takes the first enumerated candidate among them.
Identical training points are stored once, as a cell with a count per class. On data without repeated values the leave-one-out step takes about 1.5 times as long as in 0.2.0.

The fitted model no longer depends on the platform or on the order of the training rows. In 0.2.0, iris with subspace_size=(1, 2) selected three subspaces on macOS and four on Linux and Windows. CI now compares fitted models with committed results on all three.
The selection_path_ docstring said the loss was class-balanced whatever balance_classes said.

Documentation site built with MkDocs and Material for MkDocs, published to GitHub Pages: a user guide, the method and benchmark notes, an API reference generated from the docstrings, and the changelog.
Complementary selection, the new default (selection="complementary"). The ensemble is built by greedy forward selection with replacement on the candidates' out-of-fold probabilities: each step adds the subspace that most reduces the ensemble's class-balanced Brier score, n_subspaces caps the number of distinct subspaces, and the weights are vote counts. Across fourteen benchmark datasets it improves mean macro-F1 over ranking subspaces individually by 1.8 to 2.6 points, depending on the subspace sizes and the number of subspaces.
Exact leave-one-out out-of-fold probabilities from a single neighbour query per subspace, the new default cv="loo". It is faster than k-fold scoring and needs no random split.
Parameters selection, max_votes and balance_classes, and the fitted attribute selection_path_, which records the out-of-fold loss after each vote.
benchmarks/run_benchmark.py, which reproduces the fourteen-dataset benchmark in docs/benchmark.md, including an ablation of the two ingredients.
References section in the README and method note, starting with Brett Kennedy's ikNN article, which the method builds on.

Breaking: complementary selection and leave-one-out scoring are the new defaults, so SubspaceKNNClassifier() selects different subspaces and makes different predictions than in 0.1.0. On the fourteen benchmark datasets, mean macro-F1 with the default pairs moves from 0.762 to 0.780; docs/benchmark.md has the per-dataset numbers. The ikNN-style behaviour of 0.1.0 is now selection="ranked", and SubspaceKNNClassifier(selection="ranked", cv=5, max_candidates=100) reproduces the 0.1.0 configuration up to how subspace scores are averaged (next item).
Breaking: subspace scores are computed on pooled out-of-fold predictions rather than averaged over folds, so any scorer works, including probability-based ones. A splitter passed as cv must partition the samples; one that leaves samples out, such as ShuffleSplit, now raises a ValueError.
max_candidates defaults to 1000 instead of 100, which the cheaper scoring makes affordable.
weighting applies to ranked selection only; complementary selection weights by vote counts.
Plot panel titles show each subspace's weight next to its score.

SubspaceKNNClassifier: a scikit-learn compatible classifier that fits k-nearest-neighbour models on every small feature subspace, ranks them by cross-validated score, and combines the best by weighted soft or hard voting. Subspaces can have any size or a mix of sizes; wide data is handled by screening features to keep the candidate count within max_candidates.
explain, returning an Explanation per sample with one SubspaceVote per subspace, and feature_scores_ as a coarse relevance measure.
subspaceknn.plotting.plot_subspaces, drawing one-, two- and three-dimensional subspaces with decision regions where possible (optional plot extra).
Test suite covering scikit-learn's estimator contract for two configurations, behaviour, explanations, plotting, and an accuracy comparison against plain kNN on the iris, wine and breast-cancer datasets.
Project infrastructure: uv-based workflow, ruff and mypy in strict mode, CI across Python 3.10 to 3.13 and three operating systems, tag-driven PyPI release with trusted publishing, Dependabot, issue and pull request templates, contributing guide, security policy and code of conduct.