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..
DiCoFlex: Model-Agnostic Diverse Counterfactuals with Flexible Control
Authors: Oleksii Furman, Ulvi Movsum-zada, Patryk Marszałek, Maciej Zieba, Marek Śmieja
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we conduct a series of experiments that evaluate the quality of counterfactual explanations generated by various methods2. Our main objective is to understand how effectively these counterfactuals flip the predictions of a model while maintaining realism, requiring minimal changes to the input, and offering diversity among the generated alternatives. |
| Researcher Affiliation | Academia | 1Wrocław University of Science and Technology 2Faculty of Mathematics and Computer Science, Jagiellonian University, Kraków, Poland 3Doctoral School of Exact and Natural Sciences, Jagiellonian University, Kraków, Poland 4Tooploox Sp. z o.o. EMAIL, EMAIL, EMAIL, EMAIL EMAIL |
| Pseudocode | Yes | Algorithm 1 Training procedure Require: number of steps T, training examples X, classification model h( ), prior class distribution π, set of sparsity levels P, set of considered masking M, number of nearest neighbors K Initialize θ0 for t = 1 to T do |
| Open Source Code | Yes | Code available at https://github.com/ofurman/Di Co Flex |
| Open Datasets | Yes | We evaluate all methods on five well-established benchmark datasets commonly used in counterfactual explanation research: Adult [3] (income prediction), Bank Marketing [21] (customer response classification), Default [47] (credit card default risk prediction), Give Me Some Credit (GMC) [28] (financial insolvency prediction), and Lending Club [15] (loan creditworthiness classification), see Appendix B for details. |
| Dataset Splits | No | For test instances originally classified as class 0, we generate corresponding counterfactuals targeting class 1 and vice versa. Validity measures the percentage of counterfactuals that successfully change model predictions to the target class. |
| Hardware Specification | Yes | All experiments were conducted on a GPU workstation equipped with an NVIDIA RTX 4090 (24 GB VRAM) and an AMD Ryzen Threadripper PRO 5975WX CPU with 256 GB RAM. |
| Software Dependencies | No | Our experimental framework utilized Python [42] as the primary programming language, Additionally, the open-source machine learning library Py Torch [27] is used to implement Di Co Flex. |
| Experiment Setup | Yes | Each model was trained for a maximum of 1000 epochs using the Adam optimizer (lr = 10-4) and early stopping with a patience of 300 epochs based on the validation objective. Number of nearest neighbors K {8, 16, 32} Actionability penalty α {1, 10, 1000} MAF hidden features {16, 32, 64} MAF hidden layers {2, 5} Unless otherwise stated, the final configuration used K=16, α=10, 32 hidden features, and 5 hidden layers. |