Overview

Given the empirical distribution \(\widehat P\) from the training data \(\{(x_i, y_i)\}_{i \in [N]}\), we consider the following (distance-based) distributionally robust optimization formulations under the machine learning context. In general, DRO optimizes over the worst-case loss and satisfies the following structure:

\[ \min_{f \in \mathcal{F}}\max_{Q \in \mathcal{P}}\mathbb{E}_Q[\ell(f(X), Y)], \]

where \(\mathcal{P}\) is denoted as the ambiguity set. Usually, it satisfies the following structure:

\[ \mathcal{P}(d, \epsilon) = \{Q: d(Q, \widehat P) \leq \epsilon\}. \]

Here, \(d(\cdot, \cdot)\) is a notion of distance between probability measures and \(\epsilon\) captures the size of the ambiguity set.

Given each function class \(\mathcal{F}\), we classify all the models into the following cases, where each case can be further classified given each distance type \(d\).

We design our package based on the principle pipeline “Data -> Model -> Evaluation / Diagnostics” and discuss them one by one as follows:

Data Module

Synthetic Data Generation

Following the general pipeline of “Data -> Model -> Evaluation / Diagnostics”, we first integrate different kinds of synthetic data generating mechanisms into dro, including:

Python Module Function Name Description




dro.data.dataloader_classification		 classification_basic Basic classification task
classification_DN21 Following Section 3.1.1 of
"Learning Models with Uniform Performance via Distributionally Robust Optimization"
classification_SNVD20 Following Section 5.1 of
"Certifying Some Distributional Robustness with Principled Adversarial Training"
classification_LWLC Following Section 4.1 (Classification) of
"Distributionally Robust Optimization with Data Geometry"





dro.data.dataloader_regression		 regression_basic Basic regression task
regression_DN20_1 Following Section 3.1.2 of
"Learning Models with Uniform Performance via Distributionally Robust Optimization"
regression_DN20_2 Following Section 3.1.3 of
"Learning Models with Uniform Performance via Distributionally Robust Optimization"
regression_DN20_3 Following Section 3.3 of
"Learning Models with Uniform Performance via Distributionally Robust Optimization"
regression_LWLC Following Section 4.1 (Regression)
of "Distributionally Robust Optimization with Data Geometry"

Model Module

Models expose the robust empirical optimization objective from the most recent fit as model.robust_obj whenever it can be computed; otherwise, the value is None.

Linear and Kernel Models

Model and Loss Setup

For linear models, \(f(X) = \theta^{\top}X + b\). Across the linear module, the coefficient vector \(\theta = (\theta_1,\ldots, \theta_p)\) is stored as theta and the intercept \(b\) as b. Set model_type to the name in parentheses for one of the following losses.

Classification, where \(Y \in \{-1, 1\}\):

  • SVM (hinge) loss (svm): \(\ell(f(X), Y) = \max\{1 - Y f(X), 0\}\).

  • Logistic loss (logistic): \(\ell(f(X), Y) = \log(1 + \exp(-Y f(X)))\).

Regression:

  • Least absolute deviation (lad): \(\ell(f(X), Y) = |Y - f(X)|\).

  • Ordinary least squares (ols): \(\ell(f(X), Y) = (Y - f(X))^2\).

The convex linear models use CLARABEL through cvxpy by default. You can select another installed solver, such as SCS or the licensed MOSEK, when it supports the model’s formulation.

Kernel Setup

Kernelized distributionally robust regression and classification are configured through .update_kernel() and support the same four loss types. The predictor becomes \(f(X) = \sum_{i \in [N]}\alpha_i K(x, x_i)\), where \(K(\cdot,\cdot)\) is the kernel and \(\{\alpha_i\}_{i \in [N]}\) are fitted parameters.

The kernel interface follows scikit-learn:

  • metric: additive_chi2, chi2, linear, poly, polynomial, or rbf.

  • kernel_gamma: a positive gamma value, scale, or auto.

  • n_components: None for the full kernel representation, or an integer for a Nystroem approximation when \(n\) is large.

Supported DRO Methods

  • WDRO: basic Wasserstein DRO and satisficing Wasserstein DRO.

  • Standard \(f\)-DRO: KL-DRO, \(\chi^2\)-DRO, and TV-DRO.

  • Generalized \(f\)-DRO: CVaR-DRO, Marginal DRO (CVaR), and Conditional DRO (CVaR).

  • MMD-DRO.

  • Bayesian-based DRO: Bayesian-PDRO and PDRO.

  • Hybrid-distance DRO: Sinkhorn-DRO, HR-DRO, MOT-DRO, and Outlier-Robust Wasserstein DRO (OR-Wasserstein DRO).

Neural Network Models (Approximate Fitting)

The neural module implements \(\chi^2\)-DRO, CVaR-DRO, Wasserstein DRO through adversarial training, and Holistic Robust DRO. Supported architectures are linear models, vanilla MLP, AlexNet, and ResNet18. Users can also supply their own architecture through BaseNNDRO.update_user_mode(...).

See the neural-model guide for training interfaces and the Neural DRO robustness certificates guide for post-training bounds on the worst-case loss of a fitted network.

Tree-based Ensemble Models (Approximate Fitting)

The tree module supports KL-DRO, CVaR-DRO, and \(\chi^2\)-DRO with LightGBM and XGBoost.

Evaluation and Diagnostics

Some linear DRO models provide worst_distribution to inspect worst-case model performance. The evaluate function in BaseLinearDRO provides a bias-corrected loss estimate only for linear-kernel OLS and logistic models. It raises NotImplementedError for SVM, LAD, and nonlinear-kernel models; use score() for standard predictive metrics.

Reference

  • Daniel Kuhn, Soroosh Shafiee, and Wolfram Wiesemann. Distributionally robust optimization. arXiv preprint arXiv:2411.02549, 2024.