PolieDRO.jl

Build Status	Coverage	Documentation

PolieDRO is a novel analytics framework for classification and regression that harnesses the power and flexibility of data-driven distributionally robust optimization (DRO) to circumvent the need for regularization hyperparameters.

Features

• Build, solve and evaluate models in the PolieDRO framework using 3 implemented loss functions
  • Hinge loss: for classification models, used on the support vector machine (SVM)
  • Logistic loss: for classification models, used on the logistic regressor
  • Mean squared error: for regression models, used on the LASSO regressor
  • Mean absolute error: for regression models, using the MAE instead of MSE
• Create your own model by implementing custom loss functions in the PolieDRO framework

Quickstart

import Pkg
Pkg.add(url = "https://github.com/joaoqcouto/PolieDRO.jl")
using PolieDRO

## split some dataset into train and test sets
## one classification and one regression dataset as examples
## for classification problems y values are all either 1 or -1

# classification
Xtrain_class::Matrix{Float64}
ytrain_class::Vector{Float64}
Xtest_class::Matrix{Float64}
ytest_class::Vector{Float64}

# regression
Xtrain_reg::Matrix{Float64}
ytrain_reg::Vector{Float64}
Xtest_reg::Matrix{Float64}
ytest_reg::Vector{Float64}

# building the model
## classification
model_hl, predictor_hl = PolieDRO.build_model(Xtrain_class, ytrain_class, PolieDRO.hinge_loss)
model_ll, predictor_ll = PolieDRO.build_model(Xtrain_class, ytrain_class, PolieDRO.logistic_loss)

## regression
model_mse, predictor_mse = PolieDRO.build_model(Xtrain_reg, ytrain_reg, PolieDRO.mse_loss)

# solving the models
solve_model!(model_hl)
solve_model!(model_ll)
solve_model!(model_mse)

# evaluating the test sets
# classification
y_hl = predictor_hl(model_hl, Xtest_class)
y_ll = predictor_ll(model_ll, Xtest_class)

# regression
y_mse = predictor_mse(model_mse, Xtest_class)

# predictions of the models could then be compared to their expected values in ytest_class and ytest_reg

References

• GUTIERREZ, T.; VALLADÃO, D. ; PAGNONCELLI, B.. PolieDRO: a novel classification and regression framework with non-parametric data-driven regularization. Machine Learning, 04 2024