Learning Equivalence Classes of Bayesian Network Structures with GFlowNet
Authors: Michelle Liu, Zhaocheng Zhu, Olexa Bilaniuk, Emmanuel Bengio
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on both simulated and real-world datasets demonstrate that CPDAG-GFN performs competitively with established methods for learning CPDAG candidates from observational data. |
| Researcher Affiliation | Collaboration | Michelle Liu EMAIL Mila Québec AI Institute Zhaocheng Zhu EMAIL Mila Québec AI Institute Olexa Bilaniuk EMAIL Mila Québec AI Institute Emmanuel Bengio EMAIL Valence Labs |
| Pseudocode | No | The paper describes the method and its components in sections like 2.2 GFlow Net, 3.1 Heuristic Edge-Sparsity Filter, and 3.2 GFlow Net setup, but does not present a clearly labeled pseudocode or algorithm block. |
| Open Source Code | No | The paper mentions 'The DAG-GFN was run using the publicly available code from Deleu et al. (2022).' which refers to a baseline, but does not provide concrete access to the source code for CPDAG-GFN, nor does it contain a direct statement of code release for the presented methodology. |
| Open Datasets | Yes | In this section, we empirically assess CPDAG-GFN s performance against established methods by comparing the learned CPDAGs to a real-world proteomics dataset obtained from protein signaling networks (Sachs et al., 2005). |
| Dataset Splits | No | The paper mentions dataset sizes such as 'N = 854 continuous observations' for the Sachs dataset and 'observational data sizes ranging from small (100) to large (a million)' for synthetic data, but it does not provide specific details on how these datasets are partitioned into training, validation, or test splits for experimental reproduction. |
| Hardware Specification | Yes | All experiments were run on an Apple M4 Pro CPU (12 cores) with 24GB unified memory. |
| Software Dependencies | No | The paper mentions the use of relational graph neural networks (RGCN) and Simpl E score function, but it does not specify any software libraries or packages with their corresponding version numbers used for the implementation. |
| Experiment Setup | Yes | In addition, we designed our experiments to include several scenarios: different observational data sizes ranging from small (100) to large (a million) to demonstrate scalability, varying levels of noise in the data from small (0.01) to moderate (0.1), and different network complexities with expected degrees of 1d, 2d, and 3d, where d is the number of nodes, in which we set to d=10 in our experiments. For each experimental run, we sample K graphs, with K set to 100. |