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..
An Analysis of Causal Effect Estimation using Outcome Invariant Data Augmentation
Authors: Uzair Akbar, Niki Kilbertus, Hao Shen, Krikamol Muandet, Bo Dai
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate our approach with theoretical results in a linear setting for the infinite-sample case, and simulation and real-data experiments in the finite-sample case. We also present real data experiments to support our case. |
| Researcher Affiliation | Collaboration | Uzair Akbar Georgia Tech Niki Kilbertus TU Munich Helmholtz AI Hao Shen TU Munich Fortiss Gmb H Krikamol Muandet Rational Intelligence CISPA Bo Dai Georgia Tech Google Deep Mind |
| Pseudocode | No | The paper describes methods and theoretical results but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Other implementation details are provided in Appendix E, and the code to reproduce our results is publicly released at https://github.com/uzairakbar/causal-data-augmentation. |
| Open Datasets | Yes | Optical device dataset. The dataset from [24] consists of 3 3 pixel images X displayed on a laptop screen that cause voltage readings Y across a photo-diode. Colored MNIST. We evaluate on the colored MNIST dataset [41], where labels are spuriously correlated with image color during training, but this correlation is flipped at test time. |
| Dataset Splits | Yes | For the colored MNIST experiment, all CV implementations including baselines use 5-folds for a random search over an exponentially distributed regularization parameter with rate parameter of 1. Same is the case for simulation and optical device experiments, except that DA+IVL methods use a log-uniform distributed regularization parameter over [10 4, 1]. ... (i) vanilla CV with 20% samples held-out for validation (ii) level cross validation (LCV) for when Z is discrete, where hold-out data corresponding to 20% of the levels of Z for validation. |
| Hardware Specification | No | We briefly mention the hardware used to generate experimental results in the README.md file with the supplemental code. However, the results should be hardwareagnostic. |
| Software Dependencies | No | The paper mentions using stochastic gradient descent (SGD) and refers to various methods and libraries but does not specify any software versions for libraries or programming languages used (e.g., Python, PyTorch, etc.). |
| Experiment Setup | Yes | For the methods that use stochastic gradient descent (SGD), we use a learning rate of 0.01, batch size of 256 for 16 epochs. ... For the finite sample results of the linear SEM A from Example 2, by taking m = 32, k = 31 (dimension of G), σ = 0.1 and fixing τ = 0m,9 we sample a new f, ϵ and T Rm m from a standard normal distribution for each of the 32 experiments for every combination of κ and γ. ... We use the same neural architecture and parameters as [41] across all baselines, training with the IV-based objective described in the Appendix C. DA is implemented via small perturbations to hue, brightness, contrast, saturation, and translation, each parameterized by G β(2, 2). |