irithyll 7.9.1

Streaming machine learning in Rust -- gradient boosted trees, kernel methods, linear models, and composable pipelines
Documentation 
.__       .__  __  .__           .__  .__
│__│______│__│╱  │_│  │__ ___.__.│  │ │  │
│  ╲_  __ ╲  ╲   __╲  │  <   │  ││  │ │  │
│  ││  │ ╲╱  ││  │ │   Y  ╲___  ││  │_│  │__
│__││__│  │__││__│ │___│  ╱ ____││____╱____╱
                        ╲╱╲╱

Crates.io Documentation CI License MSRV GitHub stars

Streaming machine learning in Rust -- gradient boosted trees, kernel methods, linear models, and composable pipelines, all learning one sample at a time.

use irithyll::{pipe, normalizer, sgbt, StreamingLearner};

let mut model = pipe(normalizer()).learner(sgbt(50, 0.01));
model.train(&[100.0, 0.5], 42.0);
let prediction = model.predict(&[100.0, 0.5]);

Why irithyll?

  • 12+ streaming algorithms under one unified StreamingLearner trait
  • One sample at a time -- O(1) memory per model, no batches, no windows, no retraining
  • Composable pipelines -- chain preprocessors and learners with a builder API
  • Concept drift adaptation -- automatic model replacement when the data distribution shifts
  • Confidence intervals -- prediction uncertainty from RLS and conformal methods
  • Production-grade -- async streaming, SIMD acceleration, Arrow/Parquet I/O, ONNX export
  • Pure Rust -- zero unsafe, deterministic, serializable, 780+ tests

Algorithms

Every algorithm implements StreamingLearner -- train and predict with the same two-method interface.

Algorithm Type Use Case Per-Sample Cost
SGBT Gradient boosted trees General regression/classification O(n_steps * depth)
AdaptiveSGBT SGBT + LR scheduling Decaying/cycling learning rates O(n_steps * depth)
MulticlassSGBT One-vs-rest SGBT Multi-class classification O(classes * n_steps * depth)
MultiTargetSGBT Independent SGBTs Multi-output regression O(targets * n_steps * depth)
DistributionalSGBT Mean + variance SGBT Prediction uncertainty O(2 * n_steps * depth)
KRLS Kernel recursive LS Nonlinear regression (sin, exp, ...) O(budget^2)
RecursiveLeastSquares RLS with confidence Linear regression + uncertainty O(d^2)
StreamingLinearModel SGD linear model Fast linear baseline O(d)
StreamingPolynomialRegression Polynomial SGD Polynomial curve fitting O(d * degree)
GaussianNB Naive Bayes Text/categorical classification O(d * classes)
MondrianForest Random forest variant Streaming ensemble regression O(n_trees * depth)
LocallyWeightedRegression Memory-based Locally varying relationships O(window)

Preprocessing (implements StreamingPreprocessor):

Preprocessor Description
IncrementalNormalizer Welford's online standardization
OnlineFeatureSelector Streaming mutual-information feature selection
CCIPCA O(kd) streaming PCA without covariance matrices

Quick Start

cargo add irithyll

Factory Functions

The fastest way to get started -- one-liner construction for every algorithm:

use irithyll::{sgbt, rls, krls, gaussian_nb, mondrian, linear, StreamingLearner};

let mut trees  = sgbt(50, 0.01);        // 50 boosting steps, lr=0.01
let mut kernel = krls(1.0, 100, 1e-4);  // RBF gamma=1.0, budget=100
let mut bayes  = gaussian_nb();          // Gaussian Naive Bayes
let mut forest = mondrian(10);           // 10 Mondrian trees
let mut lin    = linear(0.01);           // SGD linear model, lr=0.01
let mut rls_m  = rls(0.99);             // RLS, forgetting factor=0.99

// All share the same interface
trees.train(&[1.0, 2.0], 3.0);
let pred = trees.predict(&[1.0, 2.0]);

Composable Pipelines

Chain preprocessors and learners with zero boilerplate:

use irithyll::{pipe, normalizer, sgbt, ccipca, StreamingLearner};

// Normalize → reduce to 5 components → gradient boosted trees
let mut model = pipe(normalizer())
    .pipe(ccipca(5))
    .learner(sgbt(50, 0.01));

model.train(&[100.0, 0.001, 50_000.0, 0.5, 1e-6, 42.0, 7.7, 0.3], 3.14);
let pred = model.predict(&[100.0, 0.001, 50_000.0, 0.5, 1e-6, 42.0, 7.7, 0.3]);

Kernel Methods (KRLS)

Learn nonlinear functions with automatic dictionary sparsification:

use irithyll::{krls, StreamingLearner};

let mut model = krls(1.0, 100, 1e-4);  // RBF kernel, budget=100

for i in 0..500 {
    let x = i as f64 * 0.01;
    model.train(&[x], x.sin());
}

let pred = model.predict(&[1.5708]);  // sin(pi/2) ~ 1.0

Prediction Intervals (RLS)

Get calibrated confidence intervals that narrow as data arrives:

use irithyll::{rls, StreamingLearner, RecursiveLeastSquares};

let mut model = rls(0.99);

for i in 0..1000 {
    let x = i as f64 * 0.01;
    model.train(&[x], 2.0 * x + 1.0);
}

let (pred, lo, hi) = model.predict_interval(&[5.0], 1.96);
// 95% CI: prediction is between lo and hi

Full Builder Pattern

For complete control over SGBT hyperparameters:

use irithyll::{SGBTConfig, SGBT, Sample};

let config = SGBTConfig::builder()
    .n_steps(100)
    .learning_rate(0.0125)
    .max_depth(6)
    .n_bins(64)
    .lambda(1.0)
    .grace_period(200)
    .feature_names(vec!["price".into(), "volume".into()])
    .build()
    .expect("valid config");

let mut model = SGBT::new(config);

for i in 0..500 {
    let x = i as f64 * 0.01;
    model.train_one(&Sample::new(vec![x], 2.0 * x + 1.0));
}

// TreeSHAP explanations
let shap = model.explain(&[3.0]);
if let Some(named) = model.explain_named(&[3.0]) {
    for (name, value) in &named.values {
        println!("{}: {:.4}", name, value);
    }
}

Concept Drift Detection

Automatic adaptation when the data distribution shifts:

use irithyll::{SGBTConfig, SGBT, Sample};
use irithyll::drift::adwin::Adwin;

let config = SGBTConfig::builder()
    .n_steps(50)
    .learning_rate(0.1)
    .drift_detector(Adwin::default())
    .build()
    .expect("valid config");

let mut model = SGBT::new(config);
// When drift is detected, trees are automatically replaced

Async Streaming

Tokio-native with bounded channels and concurrent prediction:

use irithyll::{SGBTConfig, Sample};
use irithyll::stream::AsyncSGBT;

#[tokio::main]
async fn main() -> Result<(), Box<dyn std::error::Error>> {
    let config = SGBTConfig::builder()
        .n_steps(30)
        .learning_rate(0.1)
        .build()?;

    let mut runner = AsyncSGBT::new(config);
    let sender = runner.sender();
    let predictor = runner.predictor();

    let handle = tokio::spawn(async move { runner.run().await });

    for i in 0..500 {
        let x = i as f64 * 0.01;
        sender.send(Sample::new(vec![x], 2.0 * x)).await?;
    }

    let pred = predictor.predict(&[3.0]);
    drop(sender);
    handle.await??;
    Ok(())
}

Python

import numpy as np
from irithyll_python import StreamingGBTConfig, StreamingGBT

config = StreamingGBTConfig().n_steps(50).learning_rate(0.1)
model = StreamingGBT(config)

for i in range(500):
    x = np.array([i * 0.01])
    model.train_one(x, 2.0 * x[0] + 1.0)

pred = model.predict(np.array([3.0]))
shap = model.explain(np.array([3.0]))

The SGBT Algorithm

The core gradient boosting engine is based on Gunasekara et al., 2024. The ensemble maintains n_steps boosting stages, each owning a streaming Hoeffding tree and a drift detector. For each sample (x, y):

  1. Compute the ensemble prediction F(x) = base + lr * sum(tree_s(x))
  2. For each boosting step, compute gradient/hessian of the loss at the residual
  3. Update the tree's histogram accumulators and evaluate splits via Hoeffding bound
  4. Feed the standardized error to the drift detector
  5. If drift is detected, replace the tree with a fresh alternate

Beyond the paper, irithyll adds EWMA leaf decay, lazy O(1) histogram decay, proactive tree replacement, and EFDT-style split re-evaluation for long-running non-stationary systems.

Architecture

irithyll/
  ensemble/        SGBT variants, config, multi-class/target, parallel, adaptive, distributional
  learners/        KRLS, RLS, Gaussian NB, Mondrian forests, linear/polynomial models
  pipeline/        Composable preprocessor + learner chains (StreamingPreprocessor trait)
  preprocessing/   IncrementalNormalizer, OnlineFeatureSelector, CCIPCA
  tree/            Hoeffding-bound streaming decision trees
  histogram/       Streaming histogram binning (uniform, quantile, k-means)
  drift/           Concept drift detectors (Page-Hinkley, ADWIN, DDM)
  loss/            Differentiable loss functions (squared, logistic, softmax, Huber)
  explain/         TreeSHAP, StreamingShap, importance drift monitoring
  stream/          Async tokio-based training runner and predictor handles
  metrics/         Online regression/classification metrics, conformal intervals, EWMA
  anomaly/         Half-space trees for streaming anomaly detection
  serde_support/   Model checkpoint/restore (JSON, bincode)

irithyll-python/   PyO3 Python bindings

Configuration

Parameter Default Description
n_steps 100 Number of boosting steps (trees in ensemble)
learning_rate 0.0125 Shrinkage factor applied to each tree output
feature_subsample_rate 0.75 Fraction of features sampled per tree
max_depth 6 Maximum depth of each streaming tree
n_bins 64 Number of histogram bins per feature
lambda 1.0 L2 regularization on leaf weights
gamma 0.0 Minimum gain required to make a split
grace_period 200 Minimum samples before evaluating splits
delta 1e-7 Hoeffding bound confidence parameter
drift_detector PageHinkley(0.005, 50.0) Drift detection algorithm for tree replacement
variant Standard Computational variant (Standard, Skip, MI)
leaf_half_life None (disabled) EWMA decay half-life for leaf statistics
max_tree_samples None (disabled) Proactive tree replacement threshold
split_reeval_interval None (disabled) Re-evaluation interval for max-depth leaves

Feature Flags

Feature Default Description
serde-json Yes JSON model serialization
serde-bincode No Compact binary serialization (bincode)
parallel No Rayon-based parallel tree training (ParallelSGBT)
simd No AVX2 histogram acceleration
kmeans-binning No K-means histogram binning strategy
arrow No Apache Arrow RecordBatch integration
parquet No Parquet file I/O
onnx No ONNX model export
neural-leaves No Experimental MLP leaf models
full No Enable all features

Examples

Run any example with cargo run --example <name>:

Example Description
basic_regression Linear regression with RMSE tracking
classification Binary classification with logistic loss
async_ingestion Tokio-native async training with concurrent prediction
custom_loss Implementing a custom loss function
drift_detection Abrupt concept drift with recovery analysis
model_checkpointing Save/restore models with prediction verification
streaming_metrics Prequential evaluation with windowed metrics
krls_nonlinear Kernel regression on sin(x) with ALD sparsification
ccipca_reduction Streaming PCA dimensionality reduction
rls_confidence RLS prediction intervals narrowing over time
pipeline_composition Normalizer + SGBT composable pipeline

Documentation

Minimum Supported Rust Version

The MSRV is 1.75. This is checked in CI and will only be raised in minor version bumps.

References

Gunasekara, N., Pfahringer, B., Gomes, H. M., & Bifet, A. (2024). Gradient boosted trees for evolving data streams. Machine Learning, 113, 3325-3352.

Lundberg, S. M., Erion, G., Chen, H., DeGrave, A., Prutkin, J. M., Nair, B., Katz, R., Himmelfarb, J., Banber, N., & Lee, S.-I. (2020). From local explanations to global understanding with explainable AI for trees. Nature Machine Intelligence, 2, 56-67.

Weng, J., Zhang, Y., & Hwang, W.-S. (2003). Candid covariance-free incremental principal component analysis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 25(8), 1034-1040.

License

Licensed under either of

at your option.

Contribution

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in this work by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.