A collection of boosting algorithms written in Rust 🦀.
Note. The algorithms depicted in above figure are aborted after 1 minutes, so some algorithms cannot reach their optimal value.
Some boosting algorithms use Gurobi optimizer,
so you must acquire a license to use this library.
If you have the license, you can use these boosting algorithms (boosters)
by specifying features = ["extended"]
in Cargo.toml
.
Note.
If you are trying to use the extended
feature without a Gurobi license,
the compilation fails.
I, a Ph.D. student, needed to implement boosting algorithms to show the effectiveness of a proposed algorithm. Some boosting algorithms for empirical risk minimization are implemented in Python 3, but three main issues exist.
- They are dedicated to boosting algorithm users, not to developers. Most algorithms are implemented by C/C++ internally, so implementing a new boosting algorithm by Python 3 results in a slow running time.
- These boosting algorithms are designed for a decision-tree weak learner even though the boosting protocol does not demand.
- There is no implementation for margin optimization boosting algorithms. Margin optimization is a better goal than empirical risk minimization in binary classification.
MiniBoosts is a crate to address the above issues. This crate provides:
- Two main traits, named Booster and WeakLearner, Some famous boosting algorithms, including AdaBoost, LPBoost, ERLPBoost, etc.
- Some weak learners, including Decision-Tree, Regression-Tree, etc.
Also, one can implement a new Booster or Weak Learner by implementing the above traits.
You can use Logger
struct to output log to .csv
file,
while printing the status like this:
See Research feature section for detail.
Currently, I implemented the following Boosters and Weak Learners.
If you invent a new boosting algorithm,
you can introduce it by implementing Booster
trait.
See cargo doc -F extended --open
for details.
BOOSTER |
FEATURE FLAG |
---|---|
AdaBoost by Freund and Schapire, 1997 |
|
AdaBoostV by Rätsch and Warmuth, 2005 |
|
SmoothBoost by Rocco A. Servedio, 2003 |
|
TotalBoost by Warmuth, Liao, and Rätsch, 2006 |
extended |
LPBoost by Demiriz, Bennett, and Shawe-Taylor, 2002 |
extended |
SoftBoost by Warmuth, Glocer, and Rätsch, 2007 |
extended |
ERLPBoost by Warmuth and Glocer, and Vishwanathan, 2008 |
extended |
CERLPBoost (Corrective ERLPBoost) by Shalev-Shwartz and Singer, 2010 |
extended |
MLPBoost by Mitsuboshi, Hatano, and Takimoto, 2022 |
extended |
GBM (Gradient Boosting Machine), by Jerome H. Friedman, 2001 |
|
GraphSepBoost (Graph Separation Boosting) by Alon, Gonen, Hazan, and Moran, 2023 |
Note. Currently, GraphSepBoost
only supports the aggregation rule
shown in Lemma 4.2 of their paper.
WEAK LEARNER |
---|
DecisionTree (Decision Tree) |
RegressionTree (Regression Tree) |
BadBaseLearner (A worst-case weak learner for LPBoost) |
GaussianNB (Gaussian Naive Bayes) |
NeuralNetwork (Neural Network, Experimental) |
-
Boosters
-
Weak Learners
- Bag of words
- TF-IDF
- RBF-Net
-
Others
- Parallelization
- LP/QP solver (This work allows you to use
extended
features without a license).
You can see the document by cargo doc --open
command.
You need to write the following line to Cargo.toml
.
miniboosts = { version = "0.3.3" }
If you want to use extended
features, such as LPBoost
, specify the option:
miniboosts = { version = "0.3.3", features = ["extended"] }
Here is a sample code:
use miniboosts::prelude::*;
fn main() {
// Set file name
let file = "/path/to/input/data.csv";
// Read the CSV file
// The column named `class` corresponds to the labels (targets).
let sample = SampleReader::new()
.file(file)
.has_header(true)
.target_feature("class")
.read()
.unwrap();
// Set tolerance parameter as `0.01`.
let tol: f64 = 0.01;
// Initialize Booster
let mut booster = AdaBoost::init(&sample)
.tolerance(tol); // Set the tolerance parameter.
// Construct `DecisionTree` Weak Learner from `DecisionTreeBuilder`.
let weak_learner = DecisionTreeBuilder::new(&sample)
.max_depth(3) // Specify the max depth (default is 2)
.criterion(Criterion::Twoing) // Choose the split rule.
.build(); // Build `DecisionTree`.
// Run the boosting algorithm
// Each booster returns a combined hypothesis.
let f = booster.run(&weak_learner);
// Get the batch prediction for all examples in `data`.
let predictions = f.predict_all(&sample);
// You can predict the `i`th instance.
let i = 0_usize;
let prediction = f.predict(&sample, i);
}
You can also save the hypothesis f: CombinedHypthesis
with the JSON
format
since it implements the Serialize/Deserialize
trait.
If you use boosting for soft margin optimization, initialize booster like this:
let n_sample = sample.shape().0;
let nu = n_sample as f64 * 0.2;
let lpboost = LPBoost::init(&sample)
.tolerance(tol)
.nu(nu); // Setting the capping parameter.
Note that the capping parameter must satisfies 1 <= nu && nu <= n_sample
.
When you invent a new boosting algorithm and write a paper, you need to compare it to previous works to show the effectiveness of your one. One way to compare the algorithms is to plot the curve for objective value or train/test loss. This crate can output a CSV file for such values in each step.
Here is an example:
use miniboosts::prelude::*;
use miniboosts::research::Logger;
use miniboosts::common::objective_functions::SoftMarginObjective;
// Define a loss function
fn zero_one_loss<H>(sample: &Sample, f: &CombinedHypothesis<H>) -> f64
where H: Classifier
{
let n_sample = sample.shape().0 as f64;
let target = sample.target();
f.predict_all(sample)
.into_iter()
.zip(target.into_iter())
.map(|(fx, &y)| if fx != y as i64 { 1.0 } else { 0.0 })
.sum::<f64>()
/ n_sample
}
fn main() {
// Read the training data
let path = "/path/to/train/data.csv";
let train = SampleReader::new()
.file(path)
.has_header(true)
.target_feature("class")
.read()
.unwrap();
// Set some parameters used later.
let n_sample = train.shape().0 as f64;
let nu = 0.01 * n_sample;
// Read the test data
let path = "/path/to/test/data.csv";
let test = SampleReader::new()
.file(path)
.has_header(true)
.target_feature("class")
.read()
.unwrap();
let booster = LPBoost::init(&train)
.tolerance(0.01)
.nu(nu);
let weak_learner = DecisionTreeBuilder::new(&train)
.max_depth(2)
.criterion(Criterion::Entropy)
.build();
// Set the objective function.
// One can use your own function by implementing ObjectiveFunction trait.
let objective = SoftMarginObjective::new(nu);
let mut logger = LoggerBuilder::new()
.booster(booster)
.weak_learner(tree)
.train_sample(&train)
.test_sample(&test)
.objective_function(objective)
.loss_function(zero_one_loss)
.time_limit_as_secs(120) // Terminate after 120 seconds
.print_every(10) // Print log every 10 rounds.
.build();
// Each line of `lpboost.csv` contains the following four information:
// Objective value, Train loss, Test loss, Time per iteration
// The returned value `f` is the combined hypothesis.
let f = logger.run("lpboost.csv");
}
Further, one can log your algorithm by implementing Research
trait.
Run cargo doc -F extended --open
to see more information.