Learning Gaussian DAG Models without Condition Number Bounds
Authors: Constantinos Costis Daskalakis, Anthimos Vardis Kandiros, Rui Yao
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We complement our theoretical findings with simulations on synthetic datasets that confirm our predictions. ... To validate our findings, we run Algorithm 1, Algorithm 2, and the algorithm in (Gao et al., 2022) on synthetic datasets generated by DAGs) and compare their performance. |
| Researcher Affiliation | Academia | 1Department of EECS, Massachusetts Institute of Technology, Cambridge, Massachusetts, United States. Correspondence to: Constantinos Daskalakis <EMAIL>, Vardis Kandiros <EMAIL>, Rui Yao <EMAIL>. |
| Pseudocode | Yes | Algorithm 1 Estimating the DAG Information Theoretically ... Algorithm 2 Efficient algorithm for recovering the topology ... Algorithm 3 Efficient algorithm for recovering the topology |
| Open Source Code | Yes | More simulations and code, including comparisons with the PC algorithm, can be found in Supplementary Materials E.2, F, and G. |
| Open Datasets | Yes | To validate our findings, we run Algorithm 1, Algorithm 2, and the algorithm in (Gao et al., 2022) on synthetic datasets generated by DAGs) and compare their performance. The samples are generated i.i.d. from a randomly generated DAG G, which is constructed as follows: (1) Draw a random permutation of {0, 1, . . . , n 1} as σ(0), . . . , σ(n 1). (2) For each node σ(j), choose σ(0), . . . , σ(j 1) as a parent of σ(j) i.i.d. with probability min( d j+2, 1 2). Then, truncate the parent set to at most d nodes. (3) For each edge i j, we assign bi,j i.i.d. uniformly sampled from [ bmax, bmin] [bmin, bmax] where 0 < bmin < bmax are tunable parameters. |
| Dataset Splits | Yes | For each combination of parameters n, d, m, the accuracy is tested by generating 45 random graphs, running all three algorithms on them, and calculating the fraction of successful recoveries of the graph. ... We set 100 trials and the figure shows how much portion of graphs are correctly identified. All the graphs are random graphs. |
| Hardware Specification | Yes | The experiments were run using Python with numpy and scikit-learn package on an 11th Gen Intel Core i711800H 2.30 GHz CPU with 16GB of memory. |
| Software Dependencies | No | The experiments were run using Python with numpy and scikit-learn package on an 11th Gen Intel Core i711800H 2.30 GHz CPU with 16GB of memory. The versions of Python, numpy, and scikit-learn are not specified. |
| Experiment Setup | Yes | Listing 6. Paremeter settings and function to do the experiments ... n = 5 # Choose any n you want d = 2 # Choose any d you want b_min = 0.5 b_max = 1 # Choose any b_max you want sigma = 1 gamma = b_min**2/2 # As Gao et al 2022 repeat = 45 # Choose the number of trials you want |