DirichletRandomForest

This repository provides Julia implementations of Random Forest algorithms based on the Dirichlet distribution.

Initialization

The core Julia scripts are located in the src folder:

• dirichlet_forest_ml.jl
• dirichlet_forest_ml_distributed.jl
• MLE_vs_MoM.jl

About q_threshold (Number of Quantiles for Splitting)

In the following function signature:

fit_dirichlet_forest!(
    forest::DirichletForest,
    X::Matrix{Float64},
    Y::Matrix{Float64},
    q_threshold::Int = 500000000,
    max_depth::Int = 10,
    min_node_size::Int = 5,
    mtry::Union{Nothing, Int} = nothing,
    optimization_method::Function = estimate_parameters_mom
)

The q_threshold parameter controls how many candidate split thresholds are considered per feature.

By default, q_threshold is set to a very large number (500000000), which causes the algorithm to use all unique values of each feature as split candidates.

This approach can be inefficient, especially when working with continuous features that have many unique values.

To improve efficiency and potentially improve generalization, you can set q_threshold to a smaller number.

For example, setting q_threshold = 4 will use only the 0.25, 0.50, and 0.75 quantiles of each feature as candidate split points.

This adjustment is particularly useful when dealing with high-cardinality continuous covariates.

Example in Julia

To see an example of how to use Dirichlet_RF in Julia, refer to the Dirichlet_RF_Example.jl file located in the examples directory.

Using from R

You can also use these Julia functions from within R. To get started, install Julia and set up the JuliaCall package in R:

if (!require("JuliaCall")) install.packages("JuliaCall")
library(JuliaCall)
install_julia()

You can see an example in the examples folder: call_to_R.R.

This part is not yet complete: The main function for calling the Julia code from R is defined in the dirichlet_forest_wrapper.R file located in the src folder. To see how to use it in practice, refer to the example_usage_dirichlet_forest.R file in the examples folder.

Code Related to the Paper: Dirichlet Random Forest for Predicting Compositional Data

For the simple and complex examples discussed in the paper, please refer to the following scripts:

• example_simple_structure.R
• example_complex_structure.R
• comparision_models
• rf_for_call_to_R.jl