Counterfactual Plans under Distributional Ambiguity

Authors: Ngoc Bui, Duy Nguyen, Viet Anh Nguyen

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In Section 5, we conduct experiments on both synthetic and real-world datasets to demonstrate the efficiency of our corrections and of our COPA framework.
Researcher Affiliation Industry Ngoc Bui, Duy Nguyen, Viet Anh Nguyen Vin AI Research, Vietnam
Pseudocode No The paper describes methods such as the COPA framework using textual descriptions (e.g., 'The COPA problem (6) can be solved efficiently under mild conditions using a projected (sub)gradient descent algorithm.'), but it does not include explicitly labeled 'Pseudocode' or 'Algorithm' blocks.
Open Source Code Yes Source code can be found at https://github.com/ngocbh/COPA.
Open Datasets Yes We use three real-world datasets: German Credit (Dua & Graff, 2017; Groemping, 2019), Small Bussiness Administration (SBA) (Li et al., 2018), and Student performance (Cortez & Silva, 2008).
Dataset Splits Yes For each present dataset D1, we train a logistic classifier Cθ0 with parameter θ0 on 80% of instances of the dataset and fix this classifier to construct counterfactual plans in whole experiment.
Hardware Specification No The paper does not provide specific details regarding the hardware used for its experiments, such as GPU models, CPU specifications, or cloud computing instance types.
Software Dependencies Yes The paper mentions using 'MOSEK' as a solver, citing 'MOSEK Optimizer API for Python 9.2.10, 2019', which includes a specific version number. It also states 'In our COPA framework, we use Adam optimizer to implement Projected Gradient Descent', and uses 'Logistic Regression' and 'three-layer MLP' for classifiers, though without version numbers for these libraries.
Experiment Setup Yes Throughout the experiments, we set the number of counterfactuals to J = 5. For Di CE, we use the default parameters recommended in the Di CE source code. The Mahalanobis correction will use the counterfactual plan obtained by the Di CE method with K = 3 and the perturbation limit is 0.1. In our COPA framework, we use Adam optimizer to implement Projected Gradient Descent and ℓ2-distance to compute perturbation cost between inputs. In this experiment, we run our COPA framework with λ1 = 2.0, λ2 = 200.0.