Differentiable Causal Discovery from Interventional Data
Authors: Philippe Brouillard, Sébastien Lachapelle, Alexandre Lacoste, Simon Lacoste-Julien, Alexandre Drouin
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We tested DCDI with Gaussian densities (DCDI-G) and with normalizing flows (DCDI-DSF) on a real-world data set and several synthetic data sets. The real-world task is a flow cytometry data set from Sachs et al. [40]. Our results, reported in Appendix C.1, show that our approach performs comparably to state-of-the-art methods. In this section, we focus on synthetic data sets, since these allow for a more systematic comparison of methods against various factors of variation (type of interventions, graph size, density, type of mechanisms). |
| Researcher Affiliation | Collaboration | Philippe Brouillard Mila, Université de Montréal Sébastien Lachapelle Mila, Université de Montréal Alexandre Lacoste Element AI Simon Lacoste-Julien Mila, Université de Montréal Canada CIFAR AI Chair Alexandre Drouin Element AI |
| Pseudocode | No | The paper describes methods and procedures (e.g., the augmented Lagrangian procedure) but does not present them in a structured pseudocode or algorithm block format. |
| Open Source Code | Yes | Our implementation is available here and additional information about the baseline methods is provided in Appendix B.4. |
| Open Datasets | Yes | The real-world task is a flow cytometry data set from Sachs et al. [40]. |
| Dataset Splits | No | Most methods have an hyperparameter controlling DAG sparsity. Although performance is sensitive to this hyperparameter, many papers do not specify how it was selected. For score-based methods (GIES, CAM and DCDI), we select it by maximizing the held-out likelihood as explained in Appendix B.5 (without using the ground truth DAG). While a held-out set is used, specific percentages or counts for train/validation splits are not provided. |
| Hardware Specification | No | The experiments were in part enabled by computational resources provided by Element AI, Calcul Quebec, Compute Canada. This statement is too general and does not provide specific hardware details like GPU models, CPU types, or memory. |
| Software Dependencies | No | The paper mentions software like "KCI-test" and general frameworks, but does not provide specific version numbers for Python, deep learning frameworks (e.g., PyTorch, TensorFlow), or other key libraries. |
| Experiment Setup | Yes | Most methods have an hyperparameter controlling DAG sparsity. Although performance is sensitive to this hyperparameter, many papers do not specify how it was selected. For score-based methods (GIES, CAM and DCDI), we select it by maximizing the held-out likelihood as explained in Appendix B.5 (without using the ground truth DAG). Each data set has 10 000 samples uniformly distributed in the different interventional settings. A total of d interventions were performed, each by sampling up to 0.1d target nodes. |