Stable Differentiable Causal Discovery

Authors: Achille Nazaret, Justin Hong, Elham Azizi, David Blei

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We then evaluate it with both observational and interventional data and in both small-scale and large-scale settings. We find that SDCD outperforms existing methods in convergence speed and accuracy, and can scale to thousands of variables.
Researcher Affiliation Academia Achille Nazaret * 1 2 Justin Hong * 1 2 Elham Azizi 1 2 3 David Blei 1 4 1Department of Computer Science, Columbia University, New York, USA 2Irving Institute for Cancer Dynamics, Columbia University, New York, USA 3Department of Biomedical Engineering, Columbia University, New York, USA 4Department of Statistics, Columbia University, New York, USA.
Pseudocode Yes Algorithm 1 SDCD Algorithm 2 DAGTrim
Open Source Code Yes Code is available at github.com/azizilab/sdcd.
Open Datasets Yes We simulate observational and interventional data for a wide range of d (number of variables)... The simulations proceed as done in Brouillard et al. (2020); Bello et al. (2022)... To further validate the results against the strongest baseline, we evaluate SDCD on the simulated data generated in Brouillard et al. (2020) (DCDI)...
Dataset Splits Yes Each stage was run for 2000 epochs with a batch size of 256, and the validation loss was computed over a held-out fraction of the training dataset (20% of the data) every 20 epochs for early stopping.
Hardware Specification Yes The training time on CPU is measured on an AMD 3960x with 4-core per method; on GPU on an AMD 3960x with 16-core and an Nvidia A5000.
Software Dependencies No The paper mentions using neural networks and an Adam optimizer, and Supplementary Table 1 implies PyTorch is used by some baselines. However, it does not provide specific version numbers for any of the key software components or libraries (e.g., PyTorch version, Python version), which are necessary for full reproducibility.
Experiment Setup Yes We fixed the hyperparameters as follows: α1 := 1e 2, β1 := 2e 4, η1 := 2e 3, τ1 := 0.2, α2 := 5e 4, β2 := 5e 3, η2 := 1e 3, γ+ := 0.005, τ2 := 0.1. ...Each stage was run for 2000 epochs with a batch size of 256...