Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Collective Counterfactual Explanations: Balancing Individual Goals and Collective Dynamics
Authors: Ahmad-Reza Ehyaei, Ali Shirali, Samira Samadi
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Conduct numerical studies to support the theoretical results and the efficacy of our method (Sec. 5). In this section, we numerically evaluate our proposed collective CE for algorithmic recourse by comparing it against six baseline approaches: Wachter Wachter et al. [2017], Growing Spheres Laugel et al. [2017], CLUE Antorán et al. [2020], FOCUS Lucic et al. [2022], C-CHVAE Pawelczyk et al. [2020], and ROAR Upadhyay et al. [2021]. We conducted 100 experiments, each with a different random seed, and computed both the modification and competition cost metrics for each run. In Fig. 4, the average results across all experiments are represented by bar plots, while the standard deviation is illustrated using error bars. |
| Researcher Affiliation | Academia | Ahmad-Reza Ehyaei Max Planck Institute for Intelligent Systems, Tübingen AI Center, Tübingen, Germany EMAIL Ali Shirali University of California, Berkeley, USA EMAIL Samira Samadi Max Planck Institute for Intelligent Systems, Tübingen AI Center, Tübingen, Germany EMAIL |
| Pseudocode | Yes | Design an efficient algorithm to solve CCE with the benefit of amortized inference (Sec. 2.4). We present a projected-gradient method to solve CCE in Algorithm 1 in the appendix. Algorithm 1 Projected-Gradient CCE Solver, Algorithm 2 Unbalanced Sinkhorn Optimal Transport, Algorithm 3 Back-and-Forth Method for Optimal Map, Algorithm 4 Entropy-based Solution for Collective Counterfactual Explanations. |
| Open Source Code | No | Our experimental code will be released after review. |
| Open Datasets | Yes | We conducted experiments on three real-world datasets commonly used in the literature to evaluate recourse methods: Adult Becker and Kohavi [1996], COMPASAngwin et al. [2016], and HELOCFICO [2018], as well as one synthetic dataset, Moons Pedregosa et al. [2011], which is also frequently utilized in illustrating algorithmic recourse. |
| Dataset Splits | Yes | Each dataset is randomly split into training (80%) and test (20%) sets. |
| Hardware Specification | No | Our experiments are lightweight and can be executed on a standard laptop. Therefore, we do not provide detailed information about the computing resources used. |
| Software Dependencies | No | We implement baseline methods using the open-source CARLA Pawelczyk et al. [2021] (Counterfactual And Recourse Library) framework in Python, which offers standardized interfaces for generating counterfactual explanations and recourse interventions. |
| Experiment Setup | Yes | Baseline hyperparameters are tuned according to the guidelines in the CARLA documentation or prior literature (see Table 1). Table 1: Hyperparameters for different methods used in experiments in the Carla Package. |