Do Finetti: On Causal Effects for Exchangeable Data
Authors: Siyuan Guo, Chi Zhang, Karthika Mohan, Ferenc Huszar, Bernhard Schölkopf
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Finally, we develop an algorithm that performs simultaneous causal discovery and effect estimation given multi-environment data. empirically validate our results in Section 5. 5 Experiments We construct synthetic datasets according to causal Pólya urn model (cf. Section 3.2) and demonstrate that Do-Finetti algorithm can estimate causal effects and graphs simultaneously. |
| Researcher Affiliation | Collaboration | 1Max Planck Institute for Intelligent Systems 2Toyota Research Institute 3 Oregon State University 4 University of Cambridge |
| Pseudocode | Yes | See Algorithm 1 in Appendix I for details of the procedure. |
| Open Source Code | Yes | We included detailed experimental setup descriptions in the appendix and also provided reproducible code in supplementary materials. |
| Open Datasets | No | We construct synthetic datasets according to causal Pólya urn model (cf. Section 3.2) The data-generating process for X Y , for example, as follows: θe Beta(α, β), ψe Beta(α, β) X Y : Xe i := Ber(θe), Y e i := Ber(ψe) Xe i |
| Dataset Splits | No | The paper describes generating synthetic data and running experiments over a varying number of environments but does not specify explicit training, validation, or test dataset splits, percentages, or methodology for partitioning the data. |
| Hardware Specification | No | The experiments can be reproduced using single laptop with CPUs with a time estimate within 5 minutes. |
| Software Dependencies | No | The paper states 'The code is building on top of Guo et al. [2023a] under license CC-BY 4.0.' but does not list specific software dependencies with version numbers (e.g., Python, PyTorch, or other libraries). |
| Experiment Setup | Yes | The data-generating process for X Y , for example, as follows: θe Beta(α, β), ψe Beta(α, β) X Y : Xe i := Ber(θe), Y e i := Ber(ψe) Xe i where denotes xor operation and Xe i , Y e i denotes variable generated at i-th position in environment e and set α = 1, β = 3. We repeat the experiment for 100 times and report the mean squared error loss between predicted and analytic solutions across varying number of environments. |