xgboost-native stable

Native XGBoost inference in pure Rust — no ONNX, no C++ runtime, runs anywhere Rust does. Reads XGBoost's native JSON format directly and implements inference and exact TreeSHAP without external dependencies.

Why

ONNX Runtime's C++ dependency is painful to cross-compile, bloats binaries, and has no support for platforms like FreeBSD. xgboost-native eliminates that entire dependency chain.

Capabilities

• Pure Rust inference — no C/C++ runtime, no FFI, no ONNX
• Exact TreeSHAP — per-feature explanations that sum to the raw margin
• Early stopping aware — respects best_iteration from trained models
• Cross-platform — compiles anywhere Rust does, including FreeBSD
• Minimal dependencies — only serde and thiserror

Usage

let model = xgboost_native::Model::load(Path::new("model.json"))?;
let prob  = model.predict(&features);
let shap  = model.shap(&features);
// sum(shap.values) + shap.bias ≈ model.raw_margin(&features)

Verification

Validated against XGBoost 3.2.0 (Python) and ort (ONNX Runtime) on held-out samples:

Metric	Tolerance
Probability	< 1e-5 vs. Python, < 1e-4 vs. ort
Raw margin	< 1e-4
SHAP values (per feature)	< 1e-3

Supported

binary:logistic objective, numeric splits, XGBoost native JSON. No multi-class, no categorical splits.