Active Bayesian Causal Inference
Authors: Christian Toth, Lars Lorch, Christian Knoll, Andreas Krause, Franz Pernkopf, Robert Peharz, Julius von Kügelgen
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through simulations, we demonstrate that our approach is more data-efficient than several baselines that only focus on learning the full causal graph. This allows us to accurately learn downstream causal queries from fewer samples while providing well-calibrated uncertainty estimates for the quantities of interest. ... We evaluate ABCI by inferring the query posterior on synthetic ground-truth SCMs using several different experiment selection strategies. ... We compare against baselines... |
| Researcher Affiliation | Academia | Christian Toth TU Graz Lars Lorch ETH Zürich Christian Knoll TU Graz Andreas Krause ETH Zürich Franz Pernkopf TU Graz Robert Peharz TU Graz Julius von Kügelgen MPI for Intelligent Systems, Tübingen University of Cambridge |
| Pseudocode | Yes | Algorithm 1: GP-Di BS-ABCI for nonlinear additive Gaussian noise models |
| Open Source Code | Yes | Code available at: https://www.github.com/chritoth/active-bayesian-causal-inference |
| Open Datasets | No | We evaluate ABCI by inferring the query posterior on synthetic ground-truth SCMs using several different experiment selection strategies. ... We sample ground truth SCMs over random scale-free graphs [6] of size d = 20, with mechanisms and noise variances drawn from our model prior in Eq. (4.4). The paper describes how it *generates* synthetic data for its simulations rather than using a pre-existing publicly available dataset. Therefore, it does not provide access information for a public dataset. |
| Dataset Splits | No | We initialise all methods with 50 observational samples, and then perform experiments with a batch size of Nt = 5. (Figure 3 caption). The paper describes an active learning setting where data is sequentially collected, not pre-split into train/validation/test sets. It does not provide specific details on how to reproduce data partitioning for these |
| Hardware Specification | No | The paper mentions running "simulations" but does not provide any specific details about the hardware used (e.g., CPU, GPU models, memory, or cloud infrastructure). |
| Software Dependencies | No | The paper cites various software libraries and tools (e.g., PyTorch, BoTorch, NetworkX, Scikit-learn, Causal Discovery Toolbox) but does not specify the version numbers for any of these dependencies, which would be necessary for full reproducibility. |
| Experiment Setup | Yes | For specific prior choices and simulation details, see Appx. D. ... In Appx. E, we provide a more detailed discussion of our GP-Di BS-ABCI implementation including computational complexity and hyperparameter settings. ... Appendix E.2: Hyperparameters for the GP priors are given by αGP = βGP = 0.5. For the Gaussian root nodes, we choose conjugate normal-inverse-gamma priors with µi = 0, λi = 0.1, αR i = 0.1, βR i = 0.1. ... For the Di BS optimization, we use Adam [41] for 1000 steps with learning rate 1e-3, followed by L-BFGS [58] until convergence. |