Native LightGBM inference in pure Rust — no C++ runtime, no FFI, runs anywhere Rust does. Parses LightGBM's model.txt text format directly and implements inference and exact path-dependent TreeSHAP without external dependencies.

Why

The official LightGBM bindings drag in a cmake build and unsafe FFI to a C++ library. lightgbm-ars eliminates that entire dependency chain — pure Rust, cross-compiles cleanly, and works on platforms the C++ runtime doesn't support.

Capabilities

• Pure Rust inference — no C/C++ runtime, no FFI, no cmake
• Exact path-dependent TreeSHAP — sum(values) + bias == raw_margin
• Numerical and categorical splits — full decision_type support including all three missing-value modes
• Linear leaves — coefficient attribution preserves SHAP additivity
• Batch inference — predict_many walks each tree once across rows, keeping the tree hot in cache
• Cross-platform — compiles anywhere Rust does, including FreeBSD
• Minimal dependencies — only thiserror

Usage

use lightgbm_ars::Model;
let m = Model::load(Path::new("model.txt"))?;
let prob  = m.predict(&features)?;
let shap  = m.shap(&features)?;        // sum(values) + bias == raw
let batch = m.predict_many(&rows)?;    // row-major

Verification

Validated against Python lightgbm.Booster on synthetic, categorical, and linear-tree fixtures:

Metric	Tolerance
Probability	< 1e-6
Raw margin	< 1e-5
SHAP additivity	< 1e-5

Supported

Binary classification, numerical and categorical splits, constant and linear leaves. No training, no multi-class (num_class != 1 is rejected at load), no round-trip save, no column-major input.

For training, multi-class, or column-major data, use lightgbm3 — it binds the C++ library at the cost of a cmake build and unsafe FFI.