DAGs with No Curl: An Efficient DAG Structure Learning Approach

Authors: Yue Yu, Tian Gao, Naiyu Yin, Qiang Ji

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental studies on benchmark datasets demonstrate that our method provides comparable accuracy but better efficiency than baseline DAG structure learning methods on both linear and generalized structural equation models, often by more than one order of magnitude.
Researcher Affiliation Collaboration 1Department of Mathematics, Lehigh University, Bethlehem, PA. 2IBM Research, Yorktown Heights, NY. 3Department of Electrical, Computer, and Systems Engineering, Rensselaer Polytechnic Institute, Troy, NY.
Pseudocode Yes Algorithm 1 DAG-No Curl algorithm
Open Source Code Yes The code will be publicly released at https: //github.com/fishmoon1234/DAG-No Curl.
Open Datasets Yes We first test our algorithm in linear synthetic datasets. We employ similar experimental setups as existing works (Zheng et al., 2018). In each experiment, a random graph G is generated by using the Erd os R enyi (ER) model or the scale-free (SF) model with k expected edges... Given A0, we take n = 1000 i.i.d. samples of X... We also test nonlinear SEM and real datasets in Section 5.5 and 5.6. ... a real-world bioinformatics dataset (Sachs et al., 2005) for the discovery of a protein signaling network...
Dataset Splits No The paper mentions generating 1000 i.i.d. samples and discusses training, but does not explicitly specify a split into training, validation, and test sets with percentages or counts.
Hardware Specification No The paper mentions "CPU time" as a metric but does not provide specific details about the hardware (e.g., GPU/CPU models, memory) used for the experiments.
Software Dependencies No The paper mentions using "L-BFGS (Liu and Nocedal, 1989)", "PyTorch (Paszke et al., 2017)", and "Adam (Kingma and Ba, 2015)", but it does not provide specific version numbers for these software components.
Experiment Setup Yes We test many sets of hyperparameters and due to the page limit the full numerical results are listed in Table 4 for ER3-Gaussian and Table 5 for ER6-Gaussian cases in the supplement. It is observed that as long as λ ≥ 10, the accuracy results are all satisfactory. Among which, λ = 10^2 and λ = (10, 10^3) are generally the best values in term of both accuracy and computational efficiency. Hence, we select them as the default values and refer them as No Curl-1 and No Curl-2, respectively. ... we use a fixed threshold value ϵ = 0.3 as suggested by NOTEARS (Zheng et al., 2018).