Decisions, Counterfactual Explanations and Strategic Behavior
Authors: Stratis Tsirtsis, Manuel Gomez Rodriguez
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on synthetic and real lending and credit card data illustrate our theoretical findings and show that the counterfactual explanations and decision policies found by our algorithms achieve higher utility than several competitive baselines. |
| Researcher Affiliation | Academia | Stratis Tsirtsis Max Planck Institute for Software Systems Kaiserslautern, Germany stsirtsis@mpi-sws.org Manuel Gomez-Rodriguez Max Planck Institute for Software Systems Kaiserslautern, Germany manuelgr@mpi-sws.org |
| Pseudocode | Yes | Algorithm 1 in Appendix C provides a pseudocode implementation of the algorithm. |
| Open Source Code | Yes | An open-source implementation can be found at https://github.com/Networks-Learning/strategic-decisions. |
| Open Datasets | Yes | We experiment with two publicly available datasets: (i) the lending dataset [37], which contains information about all accepted loan applications in Lending Club during the 2007-2018 period and (ii) the credit dataset [38], which contains information about a bank s credit card payoffs. |
| Dataset Splits | No | The paper describes the datasets and the procedure to approximate P(y|x) but does not explicitly state specific train/validation/test dataset splits, percentages, or sample counts for reproduction. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., exact GPU/CPU models, processor types, or memory amounts) used for running its experiments. |
| Software Dependencies | No | The paper does not provide specific ancillary software details with version numbers (e.g., library or solver names with versions like Python 3.8, CPLEX 12.4) needed to replicate the experiment. |
| Experiment Setup | Yes | To set the values of the cost function c(xi, xj), we use the maximum percentile shift among actionable features13, similarly as in Ustun et al. [13]. More specifically, let L be the set of actionable (numerical) features and L be the set of non-actionable (discrete-valued) features14. Then, for each pair of feature values xi, xj we define the cost function as: c(xi, xj) = α maxl L |Ql(xj,l) Ql(xi,l)| if xi,l = xj,l l L otherwise, (7) where xj,l is the value of the l-th feature for the feature value xj, Ql( ) is the CDF of the numerical feature l L and α 1 is a scaling factor. As an exception, in the credit dataset, we always set the cost c(xi, xj) between two feature values to if Ql(xj,l) < Ql(xi,l) for l {Total Overdue Counts, Total Months Overdue} considering the fact that history of overdue payments cannot be erased. In both panels, we set k = 0.05m and we repeat each experiment 20 times. |