On Data Manifolds Entailed by Structural Causal Models
Authors: Ricardo Dominguez-Olmedo, Amir-Hossein Karimi, Georgios Arvanitidis, Bernhard Schölkopf
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate the methods proposed in 5 against a variety of previously proposed counterfactual explanation methods. We open source our implementation and experiments5. We consider the following prior art: ... We present the results for CFE generation in Table 1. ... We present the results for causal algorithmic recourse in Table 2. |
| Researcher Affiliation | Academia | 1Max Planck Institute for Intelligent Systems, T ubingen, Germany 2ETH Z urich, Z urich, Switzerland 3Technical University of Denmark, Lyngby, Denmark. |
| Pseudocode | No | The paper describes methods and optimization problems using mathematical equations (e.g., Equation 11 and Equation 13), but it does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | We open source our implementation and experiments5. (with footnote 5 pointing to https://github.com/Ricardo Dominguez/data-manifolds-scms) |
| Open Datasets | Yes | We consider two real-world data sets: the COMPAS recidivism dataset (Larson et al., 2016) and the Adult demographic dataset (Kohavi & Becker, 1996), for which we assume the causal graphs presented in Nabi & Shpitser (2018). |
| Dataset Splits | No | For prediction, we train both logistic regression (LR) classifiers as well as neural network (NN) classifiers with two hidden layers. We search for counterfactuals for the negatively classified individuals in the test set. This text mentions training and a test set but no specific splits or a validation set. |
| Hardware Specification | No | The paper does not specify any particular hardware used for running the experiments. It only mentions training classifiers and performing gradient descent. |
| Software Dependencies | No | We use the versatile automatic differentiation system of JAX (Bradbury et al., 2018) to differentiate through the boundary conditions of the BVP, which we solve using a fourth order collocation algorithm with residual control similar to Kierzenka & Shampine (2001). While JAX is mentioned, a specific version number is not provided, and no other software dependencies with versions are listed. |
| Experiment Setup | No | We search for counterfactuals for the negatively classified individuals in the test set. When searching for counterfactuals, we only allow changes to real-valued features. The experimental results are averaged over five random seeds. This provides some setup details but lacks specific hyperparameters such as learning rate, batch size, and number of epochs for training or optimization. |