v0.3.4

v0.3.4 (2026-05-27)

This release is published under the Apache-2.0 License.

Bug Fixes

• Fix wrong tagging and configure semantic release (776fb6f)

---

Detailed Changes: v0.3.2...v0.3.4

v0.3.1 — Documentation site, CONTRIBUTING, and CITATION

What's new

Documentation site

• New MkDocs site (Material theme) deployed to GitHub Pages: https://delgadopanadero.com/DeepGBoost/
• API reference for all public classes, generated from NumPy docstrings via mkdocstrings
• Examples section with rendered Jupyter notebooks (quickstart, regressor, classifier, serialization)
• SEO: meta descriptions, tags, social cards, social links (LinkedIn, Twitter)
• GitHub Actions workflow for automatic deployment on push to main

Contributing & Citation

• CONTRIBUTING.md — dev setup, code style, project structure, and PR checklist
• CITATION.cff — standard citation file; GitHub shows a "Cite this repository" button on the repo page

Other

• Completed missing type annotations in NumPy docstrings across all fit / _fit_layer methods
• Images moved from doc/ to docs/img/; legacy doc/ directory removed
• .claude/settings.json — replaced bypassPermissions with an explicit allow/deny permission list

v0.3.0 — DeepGBoostMultiClassifier & benchmark improvements

What's new

New feature: DeepGBoostMultiClassifier

Native softmax multiclass classifier backed by DGBFMultiOutputModel. Unlike the OvR approach in DeepGBoostClassifier, this trains a single model with shared tree splits that jointly optimise residuals for all K classes simultaneously.

• Per-class single-output trees with SoftmaxObjective (gradient/hessian from softmax cross-entropy)
• Hessian-weighted NNLS per class per layer
• predict_proba returns calibrated probabilities via softmax; binary case handled via two-column one-hot
• Full sklearn compatibility: clone, Pipeline, pickle, get_params/set_params

Module rename: gbm/ → dgbf/

Internal module reorganisation for clarity. Direct imports from deepgboost.gbm.* are a breaking change — use deepgboost.dgbf.* instead.

Benchmark improvements

• CV std fix: CrossValidationModelTest now stores per-fold scores individually so std is computed across all folds (not collapsed per-run averages → std was always 0)
• BenchmarkGenerator fix: reads *_cross_validation_test.json correctly and strips both _bootstrap_test and _cross_validation_test suffixes from dataset names
• Ablation config: ablation experiments are now configured via the "Ablations" section in config.json and run automatically via run_experiments.py — the standalone run_ablation.py script has been removed

Other

• Ruff COM812 compliance across all source files
• Removed 6 redundant/tautological tests

v0.2.1

v0.2.1

🚀 What’s New

✨ Features & Enhancements

• Hessian Residuals 🧠: Improved the learning process by integrating second-order derivatives from the Taylor expansion (Hessian). This implements a Newton-Raphson optimization approach, showing significant gains particularly in classification tasks.

• Optimized Default Params ⚙️: After extensive iteration, we’ve updated the default parameters for both Classifier and Regressor models for better out-of-the-box performance.

📈 Benchmarks & Performance

• Rerun Benchmark Results 🏆: Confirmed! We remain the top-performing model among the big four. While DeepGBoost is more versatile than RandomForest and XGBoost, we’re still pushing to sweep every single dataset.

v0.2.0

v0.2.0

✨ New features

🎲 Layer bagging — max_features parameter

Each tree within a layer now supports feature subsampling at each split via the new max_features parameter, available on both DeepGBoostRegressor and DeepGBoostClassifier. Mirrors sklearn's DecisionTreeRegressor.max_features. Defaults to None (all features, preserving previous behaviour).

⚖️ Configurable weight_solver

New weight_solver parameter controls how the T bagged trees in each layer are combined:

• "nnls" (default) — Non-Negative Least Squares optimal weights, normalised to sum 1
• "uniform" — equal weight 1/T for every tree, equivalent to RandomForest averaging

🌲 RandomForest equivalence

With n_layers=1, weight_solver="uniform", and learning_rate=1.0 the model is mathematically equivalent to a standard RandomForest.

from deepgboost import DeepGBoostRegressor

rf = DeepGBoostRegressor(
    n_layers=1,
    weight_solver="uniform",
    max_features="sqrt",
    learning_rate=1.0,
)

🔧 Improvements

📦 Callbacks refactored from a single callback.py module into a callbacks/ package with one file per class
🔁 sklearn Pipeline compatibility verified with cross_val_score and get_params/set_params tunneling tests

🐛 Bug fixes

Fixed missing comma in callbacks/init.all that caused silent string concatenation
Fixed callback class names exported from deepgboost not matching the actual class names after refactor

v0.1.0

Hello world 👋

Initial release of DeepGBoost — a Python implementation of the Distributed Gradient Boosting Forest (DGBF) algorithm, which unifies RandomForest and GradientBoosting into a graph-structured tree ensemble capable of hierarchical representation learning.

Highlights

• DeepGBoostRegressor and DeepGBoostClassifier with a scikit-learn compatible API
• Low-level functional API via DeepGBoostBooster for full training control
• Callback system for early stopping, logging, and custom hooks
• Feature importance computation and plotting (plot_importance)
• Model serialization with pickle