RFX-Fuse

RFX-Fuse is an advanced implementation of Breiman and Cutler's concept of Random Forests, serving as a unified machine learning and similarity engine that incorporates explainable similarity natively. Designed to efficiently scale to over 25 million data points with GPU acceleration, RFX-Fuse integrates functionalities that extend beyond traditional applications.

Breiman and Cutler envisioned Random Forests not merely as ensemble predictors. Their original design encompassed a comprehensive suite of features including classification, regression, unsupervised learning, proximity-based similarity, outlier detection, missing value imputation, and visualization. Unfortunately, many of these capabilities have been overlooked in modern implementations, such as the popular scikit-learn library.

With RFX-Fuse, users can achieve equivalent accuracy and functionality to several conventional industry tools using just one or two model objects. For example, a single RFX-Fuse model can produce results comparable to up to four separate tools in the realm of time series regression while providing native explainable similarity.

Key Use Cases

The following table outlines the use cases supported by RFX-Fuse compared with conventional approaches:

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)

Key Contributions

1. Native Explainable Similarity: RFX-Fuse incorporates proximity scoring to deliver comparable outputs with established methods such as Faiss, enabling insights into model decisions through Proximity Importance explanations.

2. Imputation Quality Validation for General Tabular Data: A unique feature that assesses the realism of imputed datasets without requiring ground truth labels.

Comparison of Functionalities

RFX-Fuse consolidates numerous machine learning functionalities in a single framework, offering:

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	✓	—	—	—

Performance

RFX-Fuse demonstrates impressive performance capabilities. Benchmarks conducted using an NVIDIA RTX 3060 showcase the efficiency in terms of training time and data handling:

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

Each use case is supported with complete demonstration scripts available in the examples/ directory, illustrating practical applications of RFX-Fuse in various scenarios including recommender systems, finance explainability, and anomaly detection.

Explore RFX-Fuse for a comprehensive and efficient approach to machine learning, facilitating advanced analytics and improve model interpretability.