Trust Your $\nabla$: Gradient-based Intervention Targeting for Causal Discovery
Authors: Mateusz Olko, Michał Zając, Aleksandra Nowak, Nino Scherrer, Yashas Annadani, Stefan Bauer, Łukasz Kuciński, Piotr Miłoś
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We provide extensive experiments in simulated and real-world datasets and demonstrate that GIT performs on par with competitive baselines, surpassing them in the low-data regime. |
| Researcher Affiliation | Collaboration | Mateusz Olko1,2, Michał Zaj ac3, Aleksandra Nowak2,3,4, Nino Scherrer5 Yashas Annadani6 Stefan Bauer6 Łukasz Kuci nski2,7 Piotr Miło s2,8,7 1Warsaw University, 2IDEAS NCBR, 3Jagiellonian Univeristy, Faculty of Mathematics and Computer Science, 4Jagiellonian University, Doctoral School of Exact and Natural Sciences, 5ETH Zurich, 6Helmholtz, TU Munich, 7deepsense.ai, 8Institute of Mathematics, Polish Academy of Sciences |
| Pseudocode | Yes | Algorithm 1 ONLINE CAUSAL DISCOVERY Algorithm 2 GIT S INTERVENTION TARGET SELECTION |
| Open Source Code | No | The paper states: 'Our experiments were managed using https://neptune.ai. We thank the Neptune team for providing us access to the team version and technical support.' This indicates usage of an experiment management platform but does not explicitly state or provide a link to the open-source code for their method. |
| Open Datasets | Yes | The real-world dataset consists of alarm, asia, cancer, child, earthquake, and sachs graphs, taken from the Bn Learn repository [Scutari, 2010]. |
| Dataset Splits | No | The paper states: 'We utilize an observational dataset of size 5000. We use T = 100 rounds, in each one acquiring an interventional batch of 32 samples. We distinguish two regimes: regular, with all 100 rounds (N = 3200 interventional samples), and low, with 33 rounds (N = 1056 interventional samples).' This describes data acquisition strategy and overall sample counts, but not explicit train/validation/test splits of a fixed dataset (e.g., in percentages or specific sample counts for each split). |
| Hardware Specification | Yes | We used two hardware settings, one with GPU: a single Nvidia A100, and another one with CPUs: 12 cores of Intel Xeon E5-2697 processor. |
| Software Dependencies | No | The paper mentions using specific frameworks like ENCO and DiBS and refers to external code used for dataset creation (Lippe et al. [2022]), but it does not list specific version numbers for underlying software dependencies such as Python, PyTorch, or other libraries. |
| Experiment Setup | Yes | For experiments on ENCO framework we used exactly the same parameters as reported by Lippe et al. [2022, Appendix C.1.1]. We provide them in Table 3 for the completeness of our report. Table 3: Hyperparameters used for the ENCO framework. (lists specific parameters like Sparsity regularizer, Batch size, Learning rate model, etc.) |