Ordered Counterfactual Explanation by Mixed-Integer Linear Optimization

Authors: Kentaro Kanamori, Takuya Takagi, Ken Kobayashi, Yuichi Ike, Kento Uemura, Hiroki Arimura11564-11574

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we conducted experiments on real datasets to investigate the effectiveness and behavior of our Ord CE. All the code was implemented in Python 3.7 with scikit-learn and IBM ILOG CPLEX v12.101. All the experiments were conducted on 64-bit mac OS Catalina 10.15.6 with Intel Core i9 2.4GHz CPU and 64GB memory, and we imposed a 300 second time limit for solving.
Researcher Affiliation Collaboration 1Hokkaido University, 2Fujitsu Laboratories Ltd., 3Tokyo Institute of Technology
Pseudocode No The paper describes an "Algorithm" in numbered steps within a paragraph, but does not present it as a formal pseudocode block or algorithm environment.
Open Source Code Yes 1All the code is available at https://github.com/kelicht/ordce.
Open Datasets Yes We used four real datasets: FICO (D = 23) (FICO et al. 2018), German (D = 40), Wine Quality (D = 12), and Diabetes (D = 8) (Dua and Graff 2017) datasets
Dataset Splits No We randomly split each dataset into train (75%) and test (25%) instances, and trained ℓ2-regularized logistic regression classifiers (LR), random forest classifiers (RF) with T = 100 decision trees, and two-layer Re LU network classifiers (MLP) with T = 200 neurons, on each training dataset. The paper explicitly mentions train and test splits but does not specify a separate validation split.
Hardware Specification Yes All the experiments were conducted on 64-bit mac OS Catalina 10.15.6 with Intel Core i9 2.4GHz CPU and 64GB memory, and we imposed a 300 second time limit for solving.
Software Dependencies Yes All the code was implemented in Python 3.7 with scikit-learn and IBM ILOG CPLEX v12.101.
Experiment Setup No The paper specifies parameters for its proposed method (e.g., γ = 1.0, K = 4) and model architectures (e.g., T=100 decision trees for RF, T=200 neurons for MLP), but it does not provide common training-specific hyperparameters such as learning rate, batch size, optimizer type, or number of training epochs for the classifiers.