This package was originally authored by Allardvm and wakakusa

LightGBM.jl provides a high-performance Julia interface for Microsoft's LightGBM.

The package adds a couple of convenience features:

• Automated cross-validation
• Exhaustive grid search search procedure
• Integration with MLJ, which also provides the above via different interfaces (verified only on Julia 1.6+)

Additionally, the package automatically converts all LightGBM parameters that refer to indices (e.g. categorical_feature) from Julia's one-based indices to C's zero-based indices.

A majority of the C-interfaces are implemented. A few are known to be missing and are tracked.

All major operating systems (Windows, Linux, and Mac OS X) are supported. Julia versions 1.6+ are supported.

1. Installation
2. Example
3. MLJ

To add the package to Julia:

Pkg.add("LightGBM")

This package uses LightGBM_jll to package lightgbm binaries so it works out-of-the-box.

Running tests for the package requires the use of the LightGBM example files, download and extract the LightGBM source and set the environment variable LIGHTGBM_EXAMPLES_PATH to the root of the source installation. Then you can run the tests by simply doing

Pkg.test("LightGBM")

To skip MLJ testing when running tests, set the env var DISABLE_MLJ_TESTS to anything. (You might want to do this to get the tests to run faster)

First, download LightGBM source and untar it somewhere.

cd ~
wget https://github.com/microsoft/LightGBM/archive/v3.3.5.tar.gz
tar -xf v3.3.5.tar.gz

using LightGBM
using DelimitedFiles

LIGHTGBM_SOURCE = abspath("~/LightGBM-3.3.5")

# Load LightGBM's binary classification example.
binary_test = readdlm(joinpath(LIGHTGBM_SOURCE, "examples", "binary_classification", "binary.test"), '\t')
binary_train = readdlm(joinpath(LIGHTGBM_SOURCE, "examples", "binary_classification", "binary.train"), '\t')
X_train = binary_train[:, 2:end]
y_train = binary_train[:, 1]
X_test = binary_test[:, 2:end]
y_test = binary_test[:, 1]

# Create an estimator with the desired parameters—leave other parameters at the default values.
estimator = LGBMClassification(
    objective = "binary",
    num_iterations = 100,
    learning_rate = .1,
    early_stopping_round = 5,
    feature_fraction = .8,
    bagging_fraction = .9,
    bagging_freq = 1,
    num_leaves = 1000,
    num_class = 1,
    metric = ["auc", "binary_logloss"]
)

# Fit the estimator on the training data and return its scores for the test data.
fit!(estimator, X_train, y_train, (X_test, y_test))

# Predict arbitrary data with the estimator.
predict(estimator, X_train)

# Cross-validate using a two-fold cross-validation iterable providing training indices.
splits = (collect(1:3500), collect(3501:7000))
cv(estimator, X_train, y_train, splits)

# Exhaustive search on an iterable containing all combinations of learning_rate ∈ {.1, .2} and
# bagging_fraction ∈ {.8, .9}
params = [Dict(:learning_rate => learning_rate,
               :bagging_fraction => bagging_fraction) for
          learning_rate in (.1, .2),
          bagging_fraction in (.8, .9)]
search_cv(estimator, X_train, y_train, splits, params)

# Save and load the fitted model.
filename = pwd() * "/finished.model"
savemodel(estimator, filename)
loadmodel!(estimator, filename)

This package has an interface to MLJ. Exhaustive MLJ documentation is out of scope for here, however the main things are:

The MLJ interface models are

LightGBM.MLJInterface.LGBMClassifier
LightGBM.MLJInterface.LGBMRegressor

And these have the same interface parameters as the estimators

The interface models are generally passed to MLJBase.fit or MLJBase.machine and integrated as part of a larger MLJ pipeline. An example is provided

MLJ Is only officially supported on 1.6+ (because this is what MLJ supports). Using older versions of the MLJ package may work, but your mileage may vary.

This package uses LightGBM_jll to package lightgbm binaries. JLL packages use the Artifacts system to provide the files. If you would like to override the existing files with your own binaries, you can follow the overriding the artifacts guidance.

Please don't hesitate to add yourself when you contribute to CONTRIBUTORS.md.