Score-Based Causal Discovery of Latent Variable Causal Models
Authors: Ignavier Ng, Xinshuai Dong, Haoyue Dai, Biwei Huang, Peter Spirtes, Kun Zhang
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results validate the effectiveness of the proposed methods. We conduct experiments to validate our score-based methods, by comparing them to existing methods that support causally-related latent variables, such as FOFC (Kummerfeld & Ramsey, 2016), HUANG (Huang et al., 2022), and GIN (Xie et al., 2020). The F1 scores of skeletons are reported in Table 1, while the SHDs of MECs are given in Table 2 in the supplementary material. |
| Researcher Affiliation | Academia | 1Carnegie Mellon University 2University of California, San Diego 3Mohamed bin Zayed University of Artificial Intelligence. |
| Pseudocode | Yes | Algorithm 1 Enumerating structures under Assumption 2. Algorithm 2 Degrees of freedom under Assumption 3. Algorithm 3 Enumerating structures under Assumption 3. |
| Open Source Code | No | The paper does not include an unambiguous statement that the authors are releasing the code for the work described, nor does it provide a direct link to a source-code repository for their methodology. It mentions using 'Sci Py (Virtanen et al., 2020) and Py Torch (Paszke et al., 2019) packages', which are third-party libraries. |
| Open Datasets | No | The paper describes synthetic data generation: 'For the ground truths, we consider the 1-factor models and hierarchical structures provided in Figures 4 and 5 in Appendix C. For each structure, the nonzero elements of matrices B and C are generated uniformly at random from the interval [ 2, 0.5] [0.5, 2.0].' It does not provide access information (link, DOI, specific repository, or formal citation to an existing public dataset) for a publicly available or open dataset. |
| Dataset Splits | No | The paper mentions 'sample size T {100, 300, 1000, 3000, 10000}' for experiments and describes data generation, but it does not provide specific details on how the generated data was split into training, validation, or test sets (e.g., exact percentages, sample counts, or predefined splits). |
| Hardware Specification | No | The paper states: 'The experiments for the exact search method are conducted on 16 CPUs in parallel.' This specifies the number of CPUs but lacks specific hardware details such as CPU model, manufacturer, or clock speed. |
| Software Dependencies | No | The paper mentions using 'L-BFGS (Byrd et al., 1995) implemented through Sci Py (Virtanen et al., 2020) and Py Torch (Paszke et al., 2019) packages' and 'the Adam optimizer (Kingma & Ba, 2014)'. While it references these software components, it does not specify the exact version numbers of the SciPy, PyTorch, or Adam optimizer libraries used. |
| Experiment Setup | Yes | For our exact search method, we use L-BFGS (...) with the default hyperparameters (...). For the continuous search method, i.e., SALAD-CS, (...) each subproblem is solved using the Adam optimizer (...) with 3000 iterations. (...) we run the SALAD-CS method from 10 random initializations, and select the final solution with the best score. For FOFC, we use the implementation through the py-causal package (...) with Wishart test and significance level of 0.001. |