Efficient Neural Causal Discovery without Acyclicity Constraints
Authors: Phillip Lippe, Taco Cohen, Efstratios Gavves
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In various experimental settings, ENCO recovers graphs accurately, making less than one error on graphs with 1,000 variables in less than nine hours of computation. (Page 2, Introduction) We evaluate ENCO on structure learning on synthetic datasets for systematic comparisons and realworld datasets for benchmarking against other methods in the literature. The experiments focus on graphs with categorical variables, and experiments on continuous data are included in Appendix D.5. Our code is publicly available at https://github.com/phlippe/ENCO. (Page 6, Section 4 Experiments) |
| Researcher Affiliation | Collaboration | Phillip Lippe University of Amsterdam QUVA lab p.lippe@uva.nl Taco Cohen Qualcomm AI Research tacos@qti.qualcomm.com Efstratios Gavves University of Amsterdam QUVA lab e.gavves@uva.nl |
| Pseudocode | Yes | Finally, we give an overview over the full training loop in Algorithm 1. (Page 19, Appendix A.3) |
| Open Source Code | Yes | Our code is publicly available at https://github.com/phlippe/ENCO. (Page 6, Section 4 Experiments) |
| Open Datasets | Yes | Finally, we evaluate ENCO on causal graphs from the Bayesian Network Repository (Bn Learn) (Scutari, 2010). (Page 9, Section 4.6) In case of synthetic graphs, we follow the setup of Ke et al. (2019) and create the conditional distributions from neural networks. (Page 6, Section 4.1) |
| Dataset Splits | No | The paper mentions training data for the neural networks and provides sample sizes (e.g., “5,000 observational data points and 200 per intervention”) but does not specify a separate validation split or the methodology for creating such splits. It mentions a ‘hold-out set of graphs’ for hyperparameter search, but this is distinct from data splitting for model evaluation. |
| Hardware Specification | Yes | The computation resources deployed for all experiments are a 24-core CPU with a single NVIDIA RTX3090 GPU. (Page 12, Reproducibility Statement) |
| Software Dependencies | No | Further, all methods with neural networks used the deep learning framework Py Torch (Paszke et al., 2019) which ensures a fair run time comparison across methods. (Page 36, Section C.1.2) - PyTorch mentioned without version. For GIES, we used the implementation from the R package pcalg. (Page 35, Section C.1.2) - pcalg mentioned without version. For IGSP, we used the implementation of the python package causaldag. (Page 36, Section C.1.2) - causaldag mentioned without version. In experiments, we experienced that the Adam optimizer (Kingma & Ba, 2015) speeds up the convergence of the parameters γ and θ while not interfering with the convergence guarantees in practice. (Page 20, Appendix A.3) - Adam optimizer mentioned without version. |
| Experiment Setup | Yes | The hyperparameter search was performed on a hold-out set of graphs containing two of each graph structure. ... Table 4: Hyperparameter overview for the simulated graphs dataset experiments presented in Table 1. (Page 36, Section C.1.2 & Table 4) |