v0.2.0

What's New

Added

• Convergence overhaul: relative ELBO stopping, curvature stopping, composite convergence criteria, and configurable patience windows (#95).
• K-fold cross-validated model selection via cross_validate_components and CVConfig (#79).
• Preprocessing transforms: MissingAwareLogTransformer, MissingAwarePowerTransformer, MissingAwareWinsorizer (#82).
• Preflight data diagnostics via DataReport and check_data() (#82).
• Expose prior hyperparameters hp_va, hp_vb, hp_v as VBPCA constructor parameters (#87).
• Expose niter_broadprior on VBPCA constructor (#96).
• Expose va_init (initial broad prior value) on VBPCA constructor (#97).
• Expose xprobe_fraction for auto-generated holdout probe masks, and make_xprobe_mask utility (#98).
• Expose xprobe parameter in VBPCA.fit() for explicit probe data (#86).
• Store subspace angle in learning curves (lc["angle"]) (#90).
• GitHub issue templates for bugs, features, and documentation (#83).

Fixed

• ARD stability with missing data: clamp per-iteration Va shrinkage rate and scale ARD denominator by observed-entry fraction (#86).
• _marginal_variance crash when rmempty drops columns (#74).
• variance_ attribute now available on best model returned by select_n_components (#85).
• 'rms' added to _Metric type hint in model selection (#57).

Changed

• Skip octave-parity CI job when only irrelevant files changed (#88).
• Document RMS oscillation workaround (center data before fitting) in Known Limitations.

---

Full Changelog: v0.1.0...v0.2.0

v0.1.1 — Python 3.14 wheels

Adds pre-built wheels for Python 3.14 on Linux, macOS, and Windows. No API changes.

Fixes

• Pre-built wheels for Python 3.14 — no more source-build fallback
• __version__ now reads from package metadata (no more version drift)
• Python 3.14 added to CI test matrix

v0.1.0 — Initial release

First public release of VBPCApy: Variational Bayesian PCA for Python.

Highlights

• Full VBPCA algorithm with missing-data support (dense and sparse CSR)
• scikit-learn compatible AutoEncoder estimator with fit, transform, inverse_transform
• Automatic rank selection via ARD (Automatic Relevance Determination)
• Preprocessing utilities: centering, one-hot encoding for sparse data
• Optional plotting: scree plots, loadings bar plots, variance explained
• Parity-tested against original MATLAB/Octave implementation

Install

pip install vbpca-py