Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Differentiable Causal Discovery of Linear Non-Gaussian Acyclic Models Under Unmeasured Confounding
Authors: Yoshimitsu Morinishi, Shohei Shimizu
TMLR 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate that the proposed method can estimate causal structures, including the possibility of identifying their orientations, rather than only Markov equivalence classes, under the assumption that the data are linear and follow a multivariate generalized normal distribution. Additionally, we provide proofs of the identiļ¬ability of the parameters in ADMGs when the error terms follow a multivariate generalized normal distribution. The eļ¬ectiveness of the proposed method is validated through simulations and experiments using real-world data. |
| Researcher Affiliation | Collaboration | Yoshimitsu Morinishi EMAIL Graduate School of Data Science, Shiga University; Strategy and Consulting, Data and AI, Accenture Co. Ltd. Shohei Shimizu EMAIL SANKEN; The University of Osaka; Shiga University |
| Pseudocode | Yes | Algorithm 1: Algorithm 1: ABIC Algorithm 2: Algorithm 2: ABIC Li NGAM |
| Open Source Code | No | The paper references the code implementation of ABIC (Bhattacharya et al., 2021) available at https://gitlab.com/rbhatta8/dcd for setting up implementation and incorporating prior knowledge, but does not provide a link or explicit statement for the code of their own proposed method, ABIC Li NGAM. |
| Open Datasets | Yes | We evaluated our proposed method using a sociological data repository (https://gss.norc.org/), which has also been studied in the context of Direct Li NGAM (Shimizu et al., 2011). |
| Dataset Splits | No | Sample size n {100, 500, 1000}, Number of variables d {5, 10}, Shape parameter Ī² {1, 3, 5} (MGGD). Hence, we obtain 3 2 3 = 18 total conditions. For each condition, we run 50 trials. All simulations use Python 3.8 with NumPy/SciPy. The dataset contains 1380 samples of sociological variables, such as parental education/occupation and oļ¬spring s outcomes. |
| Hardware Specification | No | The paper states: All simulations use Python 3.8 with Num Py/Sci Py. No specific hardware details (e.g., CPU, GPU models, memory) are provided. |
| Software Dependencies | Yes | All simulations use Python 3.8 with Num Py/Sci Py |
| Experiment Setup | Yes | We used the hyperparameters recommended by ABIC (https://gitlab.com/rbhatta8/dcd) as a reference when setting up our implementation. 1. Node pairs and edge assignment. For each pair of nodes (i, j) with i < j, draw a uniform random value in [0, 1]. If this value is below a predeļ¬ned threshold for directed edges, set Xi Xj and sample the coeļ¬cient Ī“ij uniformly in [ 2.0, 0.5] [0.5, 2.0]. If this value is within the threshold for bidirected edges, set Xi Xj and assign ā¦ij = ā¦ji uniformly from [ 0.7, 0.4] [0.4, 0.7]. Otherwise, no edge is placed between (i, j). 2. Diagonal entries of ā¦. Each ā¦ii is sampled from an interval [0.4, 0.7]. To ensure ā¦is positive-deļ¬nite, we add an adjustment term proportional to P(|ā¦i, i|) plus an oļ¬set in [0.1, 0.5]. 3. Shape parameter and error terms. Let Ī² {1, 3, 5} be the shape parameter of the MGGD. Compared Methods. We compare the proposed ABIC Li NGAM to: (...) BANG (Wang & Drton, 2024), a constraint-based method that exploits higher-order moments to identify bow-free ADMG directions (5% signiļ¬cance). |