betula-cluster 0.1.1 — first public release

Fast, memory-bounded, numerically stable BETULA CF-tree clustering with a from-scratch Rust core and a scikit-learn API. No LAPACK, no SciPy at runtime.

Install: pip install betula-cluster

Highlights

• ⚡ 15–40× faster than scikit-learn at N = 1,000,000 — k-means labels 1M points in 0.20 s (vs KMeans 3.3 s, Birch 8.0 s, GaussianMixture 5.5 s).
• 🪶 Bounded memory — streaming 10M points peaks at ~57 MB, flat in N (clusters data larger than RAM); an in-core KMeans needs ~5 GB.
• 🎯 Quality at parity — k-means / GMM / Ward within ≈0.01 ARI of scikit-learn; full-covariance GMM recovers anisotropic clusters; HDBSCAN-on-CF nails non-convex shapes (ARI 1.00).
• 🔢 High-dimensional normalize=True — MNIST-784 ARI 0.04 → 0.44, beating scikit-learn (0.32).
• 🧱 Numerically stable — (n, μ, S) Welford/Chan updates, PSD by construction; no catastrophic cancellation far from the origin.

What's inside

• Heads: weighted k-means (Hamerly), GMM (diagonal & full), exact Ward-HAC, HDBSCAN-on-CF, Mapper topology; automatic k (BIC / dendrogram cut).
• Streaming: partial_fit at bounded memory, DenStream / DbStream, mergeable KLL / DDSketch quantiles.
• Data: dense f32/f64, scipy.sparse (O(nnz)), mixed numeric+categorical (k-prototypes).
• Beyond labels: predict_proba, coresets, outliers, near-duplicate pairs, representatives, drift snapshots, COP-KMeans constraints, robust (Huber) insertion.
• Engineering: scikit-learn API (Pipeline / clone / GridSearchCV), typed abi3 wheel, save/load + pickle, dependency-free CLI, reusable Rust crate.

Quality bar

153-case Python suite at 100% wrapper coverage + 129 Rust tests; clippy -D warnings + fmt clean across all feature sets; Python 3.11–3.14 (single abi3 wheel).

Wheels

abi3 (CPython 3.11+): manylinux x86_64 + aarch64, macOS x86_64 + arm64, Windows x64, + sdist.

📊 Benchmarks · 📖 Docs · 📓 Examples · 📝 Changelog