Skip to content

Latest commit

 

History

History
386 lines (325 loc) · 15.4 KB

README.md

File metadata and controls

386 lines (325 loc) · 15.4 KB

DecisionTree.jl

CI Codecov Docs Stable DOI

Julia implementation of Decision Tree (CART) and Random Forest algorithms

Created and developed by Ben Sadeghi (@bensadeghi). Now maintained by the JuliaAI organization.

Available via:

  • AutoMLPipeline.jl - create complex ML pipeline structures using simple expressions
  • CombineML.jl - a heterogeneous ensemble learning package
  • MLJ.jl - a machine learning framework for Julia
  • ScikitLearn.jl - Julia implementation of the scikit-learn API

Classification

  • pre-pruning (max depth, min leaf size)
  • post-pruning (pessimistic pruning)
  • multi-threaded bagging (random forests)
  • adaptive boosting (decision stumps), using SAMME
  • cross validation (n-fold)
  • support for ordered features (encoded as Reals or Strings)

Regression

  • pre-pruning (max depth, min leaf size)
  • multi-threaded bagging (random forests)
  • cross validation (n-fold)
  • support for numerical features

Note that regression is implied if labels/targets are of type Array{Float}

Installation

You can install DecisionTree.jl using Julia's package manager

Pkg.add("DecisionTree")

ScikitLearn.jl API

DecisionTree.jl supports the ScikitLearn.jl interface and algorithms (cross-validation, hyperparameter tuning, pipelines, etc.)

Available models: DecisionTreeClassifier, DecisionTreeRegressor, RandomForestClassifier, RandomForestRegressor, AdaBoostStumpClassifier. See each model's help (eg. ?DecisionTreeRegressor at the REPL) for more information

Classification Example

Load DecisionTree package

using DecisionTree

Separate Fisher's Iris dataset features and labels

features, labels = load_data("iris")    # also see "adult" and "digits" datasets

# the data loaded are of type Array{Any}
# cast them to concrete types for better performance
features = float.(features)
labels   = string.(labels)

Pruned Tree Classifier

# train depth-truncated classifier
model = DecisionTreeClassifier(max_depth=2)
fit!(model, features, labels)
# pretty print of the tree, to a depth of 5 nodes (optional)
print_tree(model, 5)
# apply learned model
predict(model, [5.9,3.0,5.1,1.9])
# get the probability of each label
predict_proba(model, [5.9,3.0,5.1,1.9])
println(get_classes(model)) # returns the ordering of the columns in predict_proba's output
# run n-fold cross validation over 3 CV folds
# See ScikitLearn.jl for installation instructions
using ScikitLearn.CrossValidation: cross_val_score
accuracy = cross_val_score(model, features, labels, cv=3)

Also, have a look at these classification and regression notebooks.

Native API

Classification Example

Decision Tree Classifier

# train full-tree classifier
model = build_tree(labels, features)
# prune tree: merge leaves having >= 90% combined purity (default: 100%)
model = prune_tree(model, 0.9)
# pretty print of the tree, to a depth of 5 nodes (optional)
print_tree(model, 5)
# apply learned model
apply_tree(model, [5.9,3.0,5.1,1.9])
# apply model to all the sames
preds = apply_tree(model, features)
# generate confusion matrix, along with accuracy and kappa scores
DecisionTree.confusion_matrix(labels, preds)
# get the probability of each label
apply_tree_proba(model, [5.9,3.0,5.1,1.9], ["Iris-setosa", "Iris-versicolor", "Iris-virginica"])
# run 3-fold cross validation of pruned tree,
n_folds=3
accuracy = nfoldCV_tree(labels, features, n_folds)

# set of classification parameters and respective default values
# pruning_purity: purity threshold used for post-pruning (default: 1.0, no pruning)
# max_depth: maximum depth of the decision tree (default: -1, no maximum)
# min_samples_leaf: the minimum number of samples each leaf needs to have (default: 1)
# min_samples_split: the minimum number of samples in needed for a split (default: 2)
# min_purity_increase: minimum purity needed for a split (default: 0.0)
# n_subfeatures: number of features to select at random (default: 0, keep all)
# keyword rng: the random number generator or seed to use (default Random.GLOBAL_RNG)
n_subfeatures=0; max_depth=-1; min_samples_leaf=1; min_samples_split=2
min_purity_increase=0.0; pruning_purity = 1.0; seed=3

model    =   build_tree(labels, features,
                        n_subfeatures,
                        max_depth,
                        min_samples_leaf,
                        min_samples_split,
                        min_purity_increase;
                        rng = seed)

accuracy = nfoldCV_tree(labels, features,
                        n_folds,
                        pruning_purity,
                        max_depth,
                        min_samples_leaf,
                        min_samples_split,
                        min_purity_increase;
                        verbose = true,
                        rng = seed)

Random Forest Classifier

# train random forest classifier
# using 2 random features, 10 trees, 0.5 portion of samples per tree, and a maximum tree depth of 6
model = build_forest(labels, features, 2, 10, 0.5, 6)
# apply learned model
apply_forest(model, [5.9,3.0,5.1,1.9])
# get the probability of each label
apply_forest_proba(model, [5.9,3.0,5.1,1.9], ["Iris-setosa", "Iris-versicolor", "Iris-virginica"])
# add 7 more trees
model = build_forest(model, labels, features, 2, 7, 0.5, 6)
# run 3-fold cross validation for forests, using 2 random features per split
n_folds=3; n_subfeatures=2
accuracy = nfoldCV_forest(labels, features, n_folds, n_subfeatures)

# set of classification parameters and respective default values
# n_subfeatures: number of features to consider at random per split (default: -1, sqrt(# features))
# n_trees: number of trees to train (default: 10)
# partial_sampling: fraction of samples to train each tree on (default: 0.7)
# max_depth: maximum depth of the decision trees (default: no maximum)
# min_samples_leaf: the minimum number of samples each leaf needs to have (default: 5)
# min_samples_split: the minimum number of samples in needed for a split (default: 2)
# min_purity_increase: minimum purity needed for a split (default: 0.0)
# keyword rng: the random number generator or seed to use (default Random.GLOBAL_RNG)
#              multi-threaded forests must be seeded with an `Int`
n_subfeatures=-1; n_trees=10; partial_sampling=0.7; max_depth=-1
min_samples_leaf=5; min_samples_split=2; min_purity_increase=0.0; seed=3

model    =   build_forest(labels, features,
                          n_subfeatures,
                          n_trees,
                          partial_sampling,
                          max_depth,
                          min_samples_leaf,
                          min_samples_split,
                          min_purity_increase;
                          rng = seed)

accuracy = nfoldCV_forest(labels, features,
                          n_folds,
                          n_subfeatures,
                          n_trees,
                          partial_sampling,
                          max_depth,
                          min_samples_leaf,
                          min_samples_split,
                          min_purity_increase;
                          verbose = true,
                          rng = seed)

Adaptive-Boosted Decision Stumps Classifier

# train adaptive-boosted stumps, using 7 iterations
model, coeffs = build_adaboost_stumps(labels, features, 7);
# apply learned model
apply_adaboost_stumps(model, coeffs, [5.9,3.0,5.1,1.9])
# get the probability of each label
apply_adaboost_stumps_proba(model, coeffs, [5.9,3.0,5.1,1.9], ["Iris-setosa", "Iris-versicolor", "Iris-virginica"])
# run 3-fold cross validation for boosted stumps, using 7 iterations
n_iterations=7; n_folds=3
accuracy = nfoldCV_stumps(labels, features,
                          n_folds,
                          n_iterations;
                          verbose = true)

Regression Example

n, m = 10^3, 5
features = randn(n, m)
weights = rand(-2:2, m)
labels = features * weights

Regression Tree

# train regression tree
model = build_tree(labels, features)
# apply learned model
apply_tree(model, [-0.9,3.0,5.1,1.9,0.0])
# run 3-fold cross validation, returns array of coefficients of determination (R^2)
n_folds = 3
r2 = nfoldCV_tree(labels, features, n_folds)

# set of regression parameters and respective default values
# pruning_purity: purity threshold used for post-pruning (default: 1.0, no pruning)
# max_depth: maximum depth of the decision tree (default: -1, no maximum)
# min_samples_leaf: the minimum number of samples each leaf needs to have (default: 5)
# min_samples_split: the minimum number of samples in needed for a split (default: 2)
# min_purity_increase: minimum purity needed for a split (default: 0.0)
# n_subfeatures: number of features to select at random (default: 0, keep all)
# keyword rng: the random number generator or seed to use (default Random.GLOBAL_RNG)
n_subfeatures = 0; max_depth = -1; min_samples_leaf = 5
min_samples_split = 2; min_purity_increase = 0.0; pruning_purity = 1.0 ; seed=3

model = build_tree(labels, features,
                   n_subfeatures,
                   max_depth,
                   min_samples_leaf,
                   min_samples_split,
                   min_purity_increase;
                   rng = seed)

r2 =  nfoldCV_tree(labels, features,
                   n_folds,
                   pruning_purity,
                   max_depth,
                   min_samples_leaf,
                   min_samples_split,
                   min_purity_increase;
                   verbose = true,
                   rng = seed)

Regression Random Forest

# train regression forest, using 2 random features, 10 trees,
# averaging of 5 samples per leaf, and 0.7 portion of samples per tree
model = build_forest(labels, features, 2, 10, 0.7, 5)
# apply learned model
apply_forest(model, [-0.9,3.0,5.1,1.9,0.0])
# run 3-fold cross validation on regression forest, using 2 random features per split
n_subfeatures=2; n_folds=3
r2 = nfoldCV_forest(labels, features, n_folds, n_subfeatures)

# set of regression build_forest() parameters and respective default values
# n_subfeatures: number of features to consider at random per split (default: -1, sqrt(# features))
# n_trees: number of trees to train (default: 10)
# partial_sampling: fraction of samples to train each tree on (default: 0.7)
# max_depth: maximum depth of the decision trees (default: no maximum)
# min_samples_leaf: the minimum number of samples each leaf needs to have (default: 5)
# min_samples_split: the minimum number of samples in needed for a split (default: 2)
# min_purity_increase: minimum purity needed for a split (default: 0.0)
# keyword rng: the random number generator or seed to use (default Random.GLOBAL_RNG)
#              multi-threaded forests must be seeded with an `Int`
n_subfeatures=-1; n_trees=10; partial_sampling=0.7; max_depth=-1
min_samples_leaf=5; min_samples_split=2; min_purity_increase=0.0; seed=3

model = build_forest(labels, features,
                     n_subfeatures,
                     n_trees,
                     partial_sampling,
                     max_depth,
                     min_samples_leaf,
                     min_samples_split,
                     min_purity_increase;
                     rng = seed)

r2 =  nfoldCV_forest(labels, features,
                     n_folds,
                     n_subfeatures,
                     n_trees,
                     partial_sampling,
                     max_depth,
                     min_samples_leaf,
                     min_samples_split,
                     min_purity_increase;
                     verbose = true,
                     rng = seed)

Saving Models

Models can be saved to disk and loaded back with the use of the JLD2.jl package.

using JLD2
@save "model_file.jld2" model

Note that even though features and labels of type Array{Any} are supported, it is highly recommended that data be cast to explicit types (ie with float.(), string.(), etc). This significantly improves model training and prediction execution times, and also drastically reduces the size of saved models.

MLJ.jl API

To use DecsionTree.jl models in MLJ, first ensure MLJ.jl and MLJDecisionTreeInterface.jl are both in your Julia environment. For example, to install in a fresh environment:

using Pkg
Pkg.activate("my_fresh_mlj_environment", shared=true)
Pkg.add("MLJ")
Pkg.add("MLJDecisionTreeInterface")

Detailed usage instructions are available for each model using the doc method. For example:

using MLJ
doc("DecisionTreeClassifier", pkg="DecisionTree")

Available models are: AdaBoostStumpClassifier, DecisionTreeClassifier, DecisionTreeRegressor, RandomForestClassifier, RandomForestRegressor.

Feature Importances

The following methods provide measures of feature importance for all models: impurity_importance, split_importance, permutation_importance. Query the document strings for details.

Visualization

A DecisionTree model can be visualized using the print_tree-function of its native interface (for an example see above in section 'Classification Example').

In addition, an abstraction layer using AbstractTrees.jl has been implemented with the intention to facilitate visualizations, which don't rely on any implementation details of DecisionTree. For more information have a look at the docs in src/abstract_trees.jl and the wrap-function, which creates this layer for a DecisionTree model.

Apart from this, AbstractTrees.jl brings its own implementation of print_tree.

Citing the package in publications

DOI: DOI.

BibTeX entry:

@software{ben_sadeghi_2022_7359268,
  author       = {Ben Sadeghi and
                  Poom Chiarawongse and
                  Kevin Squire and
                  Daniel C. Jones and
                  Andreas Noack and
                  Cédric St-Jean and
                  Rik Huijzer and
                  Roland Schätzle and
                  Ian Butterworth and
                  Yu-Fong Peng and
                  Anthony Blaom},
  title        = {{DecisionTree.jl - A Julia implementation of the 
                   CART Decision Tree and Random Forest algorithms}},
  month        = nov,
  year         = 2022,
  publisher    = {Zenodo},
  version      = {0.11.3},
  doi          = {10.5281/zenodo.7359268},
  url          = {https://doi.org/10.5281/zenodo.7359268}
}