Abstract: This work aims to reproduce the main findings of “Improvement-Focused Causal Recourse (ICR)“(König et al., 2023) within the field of algorithmic recourse recommendations. The authors demonstrate that acceptance-focused recourse recommendation methods, like counterfactual explanations (CE), may suggest actions that revert the model’s verdict by gaming the predictor whenever possible. To tackle this, the authors introduce ICR, which focuses on improvement by optimizing for a new target variable in their causal model. It is also demonstrated that improvement guarantees consequently translate into acceptance guarantees. We can confirm the findings of the original paper. The contribution of the current study is a more extensive assessment of the robustness and generalizability of ICR. Various techniques were employed to test the algorithm’s performance under different architectural choices, such as different classifiers or optimization methods, data and model shifts, and a new dataset. Our findings suggest that ICR is more robust than CE and causal recourse (CR).
Submission Length: Regular submission (no more than 12 pages of main content)
Changes Since Last Submission: We would like to express our gratitude towards all reviewers. Their feedback enhanced our paper significantly.
- We restructured the paper as suggested by creating a separate background section which explains CE and CR in more detail
- Taking into account the reviewers suggestions, we have two new separate sections. One dedicated to the clear explanation which experiments were reproduced and which additional robustness assessment experiments were conducted. And another separate section that summarizes the challenges faced during reproducibility, and the communication with authors. This enhanced the readability of our manuscript significantly.
- We made some minor changes, that clarify some of the reviewers questions across the whole paper.
We encourage the reviewers to contact us if they have any further inquiries.
Assigned Action Editor: ~Mingming_Gong1
Submission Number: 2243
Loading