dro: A Python Package for Distributionally Robust Optimization in Machine Learning

Jiashuo Liu^†, Tianyu Wang^†, Henry Lam, Hongseok Namkoong, Jose Blanchet
_{† equal contributions (α-β order)}

dro is a python package that implements typical DRO methods on linear loss (SVM, logistic regression, and linear regression) for supervised learning tasks. It is built based on the convex optimization solver cvxpy. The dro package supports different kinds of distance metrics $d(\cdot,\cdot)$ as well as different kinds of base models (e.g., linear regression, logistic regression, SVM, neural networks...). Furthermore, it integrates different synthetic data generating mechanisms from recent research papers.

Without specified, our DRO model is to solve the following optimization problem: $$\min_{\theta} \max_{P: P \in U} E_{(X,Y) \sim P}[\ell(\theta;(X, Y))],$$ where $U$ is the so-called ambiguity set and typically of the form $U = \{P: d(P, \hat P) \leq \epsilon\}$ and $\hat P := \frac{1}{n}\sum_{i = 1}^n \delta_{(X_i, Y_i)}$ is the empirical distribution of training samples ${(X_i, Y_i)}_{i = 1}^n$. And $\epsilon$ is the hyperparameter.

Please refer to our paper for more details.

@misc{liu2025dropythonlibrary,
      title={DRO: A Python Library for Distributionally Robust Optimization in Machine Learning}, 
      author={Jiashuo Liu and Tianyu Wang and Henry Lam and Hongseok Namkoong and Jose Blanchet},
      year={2025},
      eprint={2505.23565},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2505.23565}, 
}

Installation

(1) Install dro package

To install dro package, you can simply run:

pip install dro

And it will install all required packages.

(2) Solver selection

The convex linear DRO models use CLARABEL through cvxpy by default, so no commercial solver license is needed. You can select another installed solver when its capabilities cover your model's formulation. For example:

from dro.linear_model.chi2_dro import Chi2DRO

model = Chi2DRO(input_dim=2, solver='SCS')

You can also set model.solver = 'SCS' before calling fit(). See the CVXPY solver guide for solver capabilities and options.

MOSEK (or 'GUROBI') remains an optional choice for faster implementation. To use it, run pip install mosek, obtain a license, and set solver='MOSEK' (or solver='GUROBI') when creating a model. To prepare the license for MOSEK:

• Request license at Official Website, and then the license mosek.lic will be emailed to you.

• Put your license in your home directory as follows:

  cd
  mkdir mosek
  mv /path_to_license/mosek.lic  mosek/
  

Quick Start

A simple user example is as follows:

from dro.data.dataloader_classification import classification_basic
from dro.data.draw_utils import draw_classification
from dro.linear_model.chi2_dro import Chi2DRO

# Data generating
X, y = classification_basic(d = 2, num_samples = 100, radius = 2, visualize = True)

# Chi-square DRO 
clf_model = Chi2DRO(input_dim=2, model_type = 'logistic')
clf_model.update({'eps': 0.1})
print(clf_model.fit(X, y))

For more examples, please refer to our examples.

Documentation & APIs

As for the current released version, dro supports:

(1) Synthetic data generation

(2) Linear DRO models

The models listed below are solved by exact solvers from cvxpy.

(3) NN DRO models

The models listed below are solved by gradient descent (Pytorch).

(4) Tree-based Ensemble DRO models

The models listed below are solved by function approximation (xgboost, lightgbm).

(5) Model-based Diagnostics

In linear DRO models, we provide additional interfaces for understanding the worst-case model performance and evaluating the true model performance.

For more details, please refer to https://python-dro.org for more details!

ps: our logo is generated via GPT:)

Other Reference

[1] Learning Models with Uniform Performance via Distributionally Robust Optimization. Annals of Statistics. Annals of Statistics. 2021.

[2] Certifying Some Distributional Robustness with Principled Adversarial Training. ICLR 2018.

[3] Distributionally Robust Optimization with Data Geometry. NeurIPS 2022.