Crate rill_ml 

RillML

Lightweight, serializable online machine learning for Rust applications and streaming data.

RillML provides incremental learning primitives that can be embedded directly in native Rust applications: online statistics, preprocessors, linear/logistic regression, evaluation metrics, pipelines, progressive evaluation, drift detection, online decision-making (bandits), and optional serde-based state persistence.

Quick start

use rill_ml::{
    metrics::Mae,
    models::{LinearRegression, LinearRegressionConfig},
    optim::{Optimizer, SgdConfig},
    pipeline::RegressionPipeline,
    preprocessing::StandardScaler,
    Metric, OnlineRegressor,
};

let feature_count = 2;
let scaler = StandardScaler::new(feature_count).unwrap();
let optimizer = Optimizer::sgd(
    feature_count,
    SgdConfig { learning_rate: 0.05, l2: 0.0 },
).unwrap();
let regression = LinearRegression::new(
    feature_count,
    LinearRegressionConfig { optimizer, loss: Default::default() },
).unwrap();
let mut model = RegressionPipeline::new(scaler, regression).unwrap();
let mut mae = Mae::default();

let samples = [
    ([0.1, 0.2], 0.5),
    ([0.3, 0.8], 1.4),
    ([0.6, 0.4], 1.1),
];
for (features, target) in samples {
    let prediction = model.predict(&features).unwrap();
    mae.update(target, prediction).unwrap();
    model.learn(&features, target).unwrap();
}

Re-exports

pub use error::RillError;

pub use evaluate::BinaryClassificationSample;

pub use evaluate::RegressionSample;

pub use traits::Metric;

pub use traits::OnlineBinaryClassifier;

pub use traits::OnlineRegressor;

pub use traits::OnlineStatistic;

pub use traits::SparseClassifier;

pub use traits::SparseRegressor;

pub use traits::Transformer;

Modules

banditbandit

Online decision-making: multi-armed bandits and contextual bandits.

diagnostics

Diagnostics for online models.

drift

Drift detection and adaptation.

error

Error types for RillML.

evaluate

Progressive (prequential) evaluation.

feature_hasher

Feature hashing for dimensionality reduction.

loss

Loss functions for online models.

metrics

Online evaluation metrics.

models

Online models: baseline regressors, linear regression, logistic regression, Naive Bayes classifiers, and FTRL-Proximal for sparse features.

optim

Optimizers for online linear models.

persistence

Model state persistence via a versioned Snapshot envelope.

pipeline

Static two-segment pipelines: transformer + model.

preprocessing

Feature preprocessing transformers.

sparse

Sparse feature representation for high-dimensional data.

stats

Online univariate statistics with bounded memory.

traits

Core traits shared across RillML.