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

Use Case	RFX-Fuse	Comparable Approach
Recommender Systems	1–2 models	5 tools (FAISS + XGBoost + Shap + Isolation Forests + Custom Code)
Finance Explainability	1 model	3 tools (XGBoost + Shap + Isolation Forests)
Time Series Regression	1 model	4 tools (XGBoost + Shap + Isolation Forests + FAISS)
Imputation Validation	1 model	time series methods (general tabular: RFX-Fuse)
Anomaly Detection	1 model	3 tools (Isolation Forests + Shap + Custom Code)

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 Zenodo 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

Feature	RFX-Fuse	XGBoost	sklearn RF	FAISS
Classification	✓	✓	✓	—
Regression	✓	✓	✓	—
Unsupervised	✓	—	—	—
Overall importance	✓	✓	✓	—
Local importance (per-sample)	✓	SHAP	—	—
Proximity/similarity scoring	✓	—	—	✓
Overall proximity importance	✓	—	—	—
Local proximity importance	✓	—	—	—
Top-K similar with explanations	✓	—	—	—
Outlier detection with explanations	✓	—	—	—
Missing value imputation	✓	—	—	—

Installation

From PyPI

pip install rfx-fuse

CPU-only version (pip install rfx-fuse-cpu) coming soon.

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
pip install -e . --config-settings=cmake.args=-DRFX_CPU_ONLY=ON

Prerequisites

• CMake 3.12+
• Python 3.8+
• C++ compiler with C++17 support (GCC 7+, Clang 5+)
• OpenMP (usually included with compiler)
• CUDA toolkit 12.8+ (for GPU acceleration)

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:

Use Case	Demo Script	Description
Recommender Systems	examples/recommender_system/demo_recommender_system.py	MovieLens 25M: similarity retrieval + ranking with explanations
Finance Explainability	examples/classification/demo_loan_classification.py	Loan default prediction with 4-type explainability
Time Series Regression	examples/time_series/demo_time_series.py	Bike sharing: prediction + outlier detection
Imputation Validation	examples/data_imputation/demo_imputation.py	Rank imputation methods without ground truth
Anomaly Detection	examples/anomaly_detection/demo_anomaly_detection.py	Breiman-Cutler outlier detection

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 Zenodo paper.

Recommender System Stage 2 Part 1:

Explanations available in Zenodo paper.

Recommender System Stage 2 Part 2:

Explanations available in Zenodo paper.

Recommender System Stage 2 Part 3:

Explanations available in Zenodo paper.

View Code →

---

Use Case 2: Finance Explainability

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

Explanations available in Zenodo paper.

View Code →

---

Use Case 3: Time Series Regression

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

Explanations available in Zenodo 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 Zenodo paper.

View Code →

---

Use Case 5: Anomaly Detection

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

Explanations available in Zenodo 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

Use Case	Train Size	Features	Trees	Training Time
Recommender (Unsup)	59,047 (×2)	23	1,000	1,254s
Recommender (Sup)	47,237	21	1,000	120s
Finance Classification	46,396	15	500	69s
Bike Regression	5,725	4	1,000	24s
Imputation Validation	3,000	12	100	3.6s
Anomaly Detection	15,000	8	100	112s

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.