Changelog

The canonical release notes live in CHANGELOG.md at the repository root. Keep that file synchronized with the package version, PyPI release, GitHub tag, and Read the Docs build.

Recent releases

1.1.19 — 2026-07-23

  • Added a reproducible OpenML-CC18 benchmark workflow and static dashboard for 36 classification datasets using official task-defined splits.

  • Added rich, on-demand Dash data profiling with role-aware previews, quality findings, missingness, distributions, correlations, and target associations.

  • Improved RPTE representation handling by retaining higher-order patterns as direct terms and consolidating equivalent leaf and pattern terms into one canonical component with auditable aliases.

  • Refactored the classifier implementation into focused internal modules while preserving the public estimator API, sklearn behavior, serialization, and package imports.

  • Added consistent Default, Guided, and Advanced model configuration modes and a dedicated Results complexity view in the Dash interface.

  • Aligned EBM inspection complexity with the other interpretable models by expanding active interaction score cells by source-feature arity at both model and instance levels.

1.1.18 — 2026-07-19

  • Added the optional RPTE downstream model for higher-order relationships with adaptive sequential or bounded-lookahead tree growth.

  • Added interaction relaxation through the first child split, with held-out acceptance checks for jointly useful features.

  • Added interpretable RPTE tree, rule-table, direct-term provenance, complexity, instance inspection, and structured serialization views.

  • Added a Dash-based Governance Studio while retaining the lightweight Streamlit interface.

  • Expanded benchmark analysis to 100 diverse datasets: 50 real-world and 50 synthetic datasets.

1.1.17 — 2026-07-07

  • Added a native L=3 mining hot path for bounded three-item pattern workloads.

  • Added max_mining_seconds and get_mining_audit_log() for explicit mining-stage time budgets and compact attempt-level audit trails.

  • Optimized L3 mining with tighter same-feature checks, pair reuse, coverage de-duplication, and gated pair-cache pruning.

  • Improved native timeout behavior so miners can return partial results where possible.

  • Added focused fuzz, timeout, L3 hot-path, and zero-variance tests for the new mining paths and edge cases.

  • Added lr_solver support for built-in downstream saga LogisticRegression and sgd SGDClassifier paths while preserving the existing default solver behavior.

  • Added scalability dashboard solver scenarios for the new saga and SGD downstream paths.

  • Added a privacy-sanitized reproducibility/SBOM manifest option for the scalability and benchmark dashboard.

  • Updated README and Sphinx guidance for choosing downstream solvers, and updated serialization for the new built-in SGD solver state.

1.1.16 — 2026-06-30

  • Added the optional HUGIML LLM assistant with Streamlit and CLI entry points, built-in demo datasets, deterministic fast mode, and optional Ollama-backed thinking mode.

  • Improved augmented-pair adaptive path reuse so repeated tuning and scoring flows can reuse prepared path state more efficiently.

  • Expanded native HUIM pruning behavior around LIU, LA, and gated EUCS, with documented EUCS environment controls.

  • Refreshed Sphinx documentation, Read the Docs navigation, and static benchmark/scalability outputs for the new release.

1.1.15 — 2026-06-27

  • Added a deterministic constant-prior fallback for patterns_only fits when no HUG patterns are mined.

  • Persisted fallback state in the versioned model format so fallback predictions round-trip through save/load.

  • Strengthened interaction_relaxed_mining by using joint survivor-pair evidence during native external-utility admission.

  • Refreshed interpretable benchmark grids, including compact EBM settings and a registered RuleFit grid.

1.1.14 — 2026-06-25

  • Reduced memory pressure for adaptive-binning workflows by adding deterministic stratified sampling for native bin-count selection.

  • Preserved float32 numeric inputs through the native L1 preparation path where possible, avoiding unnecessary widening copies while keeping C-contiguous arrays for native execution.

  • Exposed the sampling option in Workbench Advanced configuration and refreshed the Governance Studio static preview to match the new control.

1.1.13 — 2026-06-24

  • Improved fit and predict runtime through faster categorical processing, optional categorical-code reuse, and consistent zero-variance column exclusion.

  • Added centralized HUGIML and benchmark hyperparameter grid configuration used by classifier tuning, the benchmark runner, and the dashboard Workbench.

  • Added experiments/ reproducibility workflows for regenerating benchmark and scalability dashboard outputs.

  • Updated tuning defaults so candidate sweeps use leaner execution while the returned refit model remains audit-ready unless production mode is explicitly requested.

1.1.12 — 2026-06-20

  • Added interaction-information source selection for augmented-pair features, including augmented_pair_mode, aug_feature_size, and ii_partner_size.

  • Added interaction_relaxed_mining for survivor-led native mining without pair-operator feature columns.

  • Added pair-aware adaptive-binning metadata and survivor-led audit fields for governance review.

  • Kept augmented_pair_max_features as a v1.1.11-compatible alias for the new augmented-pair source budgets.

1.1.11 — 2026-06-17

  • Added expanded Governance Studio evidence views for dashboard review, including adaptive binning, augmented-pair traceability, feature-family review, pattern coverage, cross-validation monitoring, and export-ready governance summaries.

  • Added binary feature conversion: two-value numeric columns are treated as categorical indicators before HUG pattern mining and downstream feature preparation.

1.1.10 — 2026-06-12

  • Strengthened build and import reliability for the native backend with clearer install diagnostics, a public check_native() verification helper, cleaned package exports, and preserved classifier naming compatibility.

  • Upgraded Governance Studio into a richer workbench experience for dataset setup, candidate model comparison, drill-down review, workflow guidance, demo data, and complete governance evidence views.

1.1.9 — 2026-06-10

  • Added HUGIMLClassifier as the primary public class name while keeping HUGIMLClassifierNative backward-compatible.

  • Added execution_mode="audit"|"production". Audit is the default; production keeps prediction and save/load state while omitting large training-review artifacts.

  • Added clear guidance from audit/governance methods when production-mode models do not retain full traceability data.

  • Reduced memory pressure in strict and hybrid modes through earlier original-feature selection, compact native code storage, CSR-returning matrix paths, and conditional dense/CSR downstream handling.

  • Improved fixed-B numeric handling so clean numeric columns remain numeric and only training columns with missing or infinite values use missing-aware bin labels.

  • Optimized hybrid prediction so selected original columns are prepared directly.

1.1.8 — 2026-06-08

  • Added fast tuning for eligible adaptive-binning hyperparameter searches, reducing repeated mining work during cross-validation while preserving ordinary tuning behavior for unsupported grids.

  • Added HUGIML Governance Studio, a Streamlit dashboard for validation, representation audit, pattern inventory, case review, configuration comparison, pruning analysis, monitoring, and report-oriented review workflows.

  • Improved feature_mode="original_plus_interactions" so higher-order pattern selection is based on structural pattern order rather than parsing display labels.

  • Expanded benchmark and dashboard runner utilities for more consistent validation-fold comparisons.

1.1.7 — 2026-06-05

  • Added a native L=2 mining hot path for common two-item pattern workloads.

  • Expanded augmented-pair features with two additional operations: pair sums and signed differences.

  • Revised the benchmark baseline notebook and exported HTML with the updated baseline run and current HUGIML grid settings.

  • Preserved the v1.1.6 augmented-feature APIs, feature-lineage metadata, and serialization compatibility while extending pair-operation coverage.

1.1.6 — 2026-06-04

  • Added native augmented-pair features for L > 1 adaptive-binning models, including product and absolute-difference transforms.

  • Added strict global topK budgeting across original features, HUG patterns, and augmented-pair features.

  • Added raw-scale augmented-pair interpretation metadata, downstream feature composition summaries, and hybrid-model explainability report fields.

1.1.5 — 2026-06-01

  • Reduced native transaction memory use by storing compact item ids with shared item-level utility lookup instead of repeating utility values in every materialized transaction entry.

  • Integrated adaptive binning into the fused native L=1 execution path, avoiding an intermediate binned-matrix materialization for common adaptive workflows.

  • Parallelized native adaptive bin selection and bin-code application; n_jobs is now applied before adaptive preprocessing when native support is available.

  • Improved large-data stability with clearer native memory and timeout errors, plus safer fallback behavior under memory pressure.

1.1.4 — 2026-05-31

  • Added the native L1 hot path for L=1 fits, fusing transaction preparation, single-item pattern mining, information-gain filtering, top-K retention, and sparse matrix construction in the C++ path.

  • Moved adaptive binning selection into the C++ backend with supervised information-gain scoring and elbow-style stopping.

  • Preserved Python-side adaptive-binning metadata including per_feature_b_, _bin_edges_, and ig_scores_ for inspection and serialization.

1.1.3 — 2026-05-29

  • Optimized the native mining path by pushing the effective topK budget into mining, closer to the original Java implementation.

  • Added performance and memory improvements, including row-stripe chunked transaction construction.

  • Enforced structured mining constraints exactly and kept EUCS pruning disabled by default for predictable compound-pattern behavior.

  • Revised domain-specific notebook examples.

1.1.2 — 2026-05-27

  • Added feature_mode to HUGIMLClassifierNative.

  • Supported patterns_only, original_plus_patterns, and original_plus_interactions downstream representations.

  • Preserved pattern-space behavior for transform(), fit_transform(), get_hug_features(), and get_pattern_info().

  • Updated serialization, summaries, feature importances, and tests for hybrid feature modes.

1.1.1 — 2026-05-26

  • Fixed compound-pattern information-gain handling so L2+ patterns survive positive G after item-list release.

  • Added notebooks and examples.

1.1.0 — 2026-05-23

  • Added adaptive binning and supervised per-feature B selection.

  • Added native missing-value handling for numerical NaN and infinite values.

  • Added plotting, pruning, interpretability metrics, multiclass/imbalance helpers, benchmarks, and expanded documentation assets.