Markov Equivalence and Consistency in Differentiable Structure Learning
Authors: Chang Deng, Kevin Bello, Pradeep Ravikumar, Bryon Aragam
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct experiments to evaluate the advantages of using a likelihood-based score. The findings from these experiments are detailed in in Section 6. The empirical results support our theory: The likelihood-based score is robust and scale invariant. We also release open-source code to facilitate the implementation and reproduction of our results. |
| Researcher Affiliation | Academia | Chang Deng Kevin Bello , Pradeep Ravikumar Bryon Aragam Booth School of Business, University of Chicago, Chicago, IL 60637 Machine Learning Department, Carnegie Mellon University, Pittsburgh, PA 15213 |
| Pseudocode | Yes | Algorithm 1: Full Implementation Input: Sample covariance bΣ, decay factor γ (0, 1), and λ, δ, initial point (Bin, Ωin), initial loss ℓin (typically very large) // initial point (Bin, Ωin) is obtained from NOTEARS or DAGMA Output: (Best, Ωest) 1 while True do 2 Solve LOGLL-NOTEARS or LOGLL-DAGMA with input (Bin, Ωin), and get output (Bout, Ωout) 3 Calculate ℓn(Bout, Ωout) 4 if ℓin > ℓn(Bout, Ωout) then 5 ℓin ℓn(Bout, Ωout) 8 (Bin, Ωin) (Bout, Ωout) 10 return (Bin, Ωin) |
| Open Source Code | Yes | Opensource code is available at https://github.com/duntrain/dagrad. ... We also release open-source code to facilitate the implementation and reproduction of our results. |
| Open Datasets | No | We generated random datasets X Rn p by sampling rows i.i.d. from the models described above. For each simulation, we produced datasets with n samples across graphs with p nodes. |
| Dataset Splits | No | The paper describes generating random datasets but does not explicitly provide details about training, validation, or test splits. It mentions simulation parameters like 'n samples' and 'p nodes' for dataset generation. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running the experiments (e.g., specific GPU/CPU models, memory amounts, or cloud instances). |
| Software Dependencies | No | The implementation is based on the py-tetrad package, available at https://github.com/cmu-phil/py-tetrad. ... It is implemented using Python and TensorFlow. ... The code is implemented in Python and is available https://github.com/Scriddie/Varsortability. ... We replaced LBFGS-B with ADAM [27]... (No specific version numbers for these software components are provided). |
| Experiment Setup | Yes | After running the algorithms, a post-processing threshold of 0.3 was applied to the estimated matrix Best to prune small values, following the methodology in [73, 74]. ... In Algorithm 1, we detail the complete implementation of LOGLL(-NOTEARS/DAGMA)-SAMPLE. ... Empirically, we find that setting γ = 0.8, λ = 0.4, and δ = 0.2 usually serves as a good choice for the parameters in our optimization procedures. |