RFX-Fuse: Breiman and Cutler's Random Forests as a Forest Unified Learning and Similarity Engine - Extended with Native Explainable Similarity

RFX-Fuse (Random Forests X [X=compression] — Forest Unified Learning and Similarity Engine) delivers Breiman and Cutler's complete vision for Random Forests as a Forests Unified Machine Learning and Similarity Engine with native GPU/CPU support.

Breiman and Cutler designed Random Forests as more than an ensemble predictor. Their original implementation from the early 2000s included classification, regression, unsupervised learning, proximity-based similarity, outlier detection, missing value imputation, and visualization. Modern libraries like scikit-learn's random forests implementation (2010-2011) skipped many of these features.

These capabilities enable it to be a unified learning and similarity engine. With just 1-2 model objects, we can achieve comparable accuracy and output to 3-5 main industery tools. For example, 1 model has comparable output to 4 separate tools for Time Series Regression + native explainable similarity. 1 model = 1 set of trees grown once.

Key Use Cases

Novel Contributions

1. Native Explainable Similarity: Breiman and Cutler's original similarity scoring via proximities enable comparable output with Faiss for NDCG + HR on retrieval. Proximity Importance gives the why.

Explanations available in arXiv paper.

2. Imputation Quality Validation for General Tabular Data — Rank imputation methods by how "real" the imputed data looks, without ground truth labels.

Comparable Tools Functionality Comparison

Installation

Pre-built Wheels (Recommended)

Pre-built binary wheels are available on PyPI for Python 3.11, 3.12, and 3.13. No compiler, CMake, or CUDA toolkit needed for CPU builds.

GPU (requires CUDA 12.4+ toolkit installed on your system)

pip install rfx-fuse

CPU-only

pip install rfx-fuse-cpu

Available Wheels

Note: GPU wheels require the CUDA Toolkit 12.4+ to be installed on your system. CPU wheels have no external dependencies beyond Python and NumPy.

From Source (GPU)

git clone https://github.com/chriskuchar/RFX-Fuse.git
cd RFX-Fuse
pip install -e .

From Source (CPU-only)

git clone https://github.com/chriskuchar/RFX-Fuse.git
cd RFX-Fuse
RFX_CPU_ONLY=1 pip install -e .

Prerequisites (source builds only)

• Python 3.9+
• CMake 3.18+
• C++ compiler with C++17 support (GCC 7+, Clang 5+, MSVC 2019+)
• OpenMP (usually included with compiler)
• CUDA toolkit 12.4+ (GPU builds only)

Verify Installation

import RFXFuse as rfx
print(f"RFX-Fuse version: {rfx.__version__}")
print(f"CUDA enabled: {rfx.__cuda_enabled__}")

Examples

Each use case has a complete demonstration script in the examples/ folder:

Run an example:

cd examples/time_series
python demo_time_series.py

Industry Use Cases

Use Case 1: Recommender Systems

RFX-Fuse Unsupervised for retrieval + RFX-Fuse Supervised for re-ranking on MovieLens 25M.

Recommender System Stage 1:

Explanations available in arXiv paper.

Recommender System Stage 2 Part 1:

Explanations available in arXiv paper.

Recommender System Stage 2 Part 2:

Explanations available in arXiv paper.

Recommender System Stage 2 Part 3:

Explanations available in arXiv paper.

View Code →

Use Case 2: Finance Explainability

Single classifier provides regulatory-compliant explanations (ECOA, GDPR, Fair Lending).

Explanations available in arXiv paper.

View Code →

Use Case 3: Time Series Regression

RFX-Fuse Regressor on UCI Bike Sharing dataset with full explainability.

Explanations available in arXiv paper.

View Code →

Use Case 4: Imputation Quality Validation

Novel capability for general tabular data. Rank imputation methods by how "real" the imputed data looks.

Explanations available in arXiv paper.

View Code →

Use Case 5: Anomaly Detection

Breiman-Cutler method: train on clean data, anomalies have high P(synthetic).

Explanations available in arXiv paper.

View Code →

API Reference

For complete API documentation with all parameters, methods, and examples, see docs/API.md.

Performance

GPU Benchmarks

Environment: NVIDIA RTX 3060 (12GB), AMD Ryzen 7 5800X, 32GB RAM

Training times include predictions, similarity scoring, proximity importance, local importance, and all explainability features where applicable.

CPU Benchmarks

Coming soon.

Methodology

For detailed methodology, see:

• arXiv: https://arxiv.org/html/2603.13234v1

Citation

@article{kuchar2026rfxfuse,
  author       = {Kuchar, Chris},
  title        = {RFX-Fuse: Breiman and Cutler's Unified ML Engine + Native Explainable Similarity},
  year         = {2026},
  journal      = {arXiv preprint arXiv:2511.19493},
  url          = {https://arxiv.org/html/2603.13234v1}
}

Acknowledgments

This work aims to implement the full unified learning and similarity engine Dr. Leo Breiman and Dr. Cutler created when they made their Fortran/Java implementation in the early 2000s.

Special thanks to Dr. Adele Cutler for generously sharing original Breiman-Cutler Random Forest source materials, which made this faithful restoration and extension possible.

Work in Progress

• Multi-class classification support

Previous Work

• This is the successor to https://github.com/chriskuchar/RFX.

License

MIT License - see LICENSE for details.