Structural Agnostic Modeling: Adversarial Learning of Causal Graphs
Authors: Diviyan Kalainathan, Olivier Goudet, Isabelle Guyon, David Lopez-Paz, Michèle Sebag
JMLR 2022 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | SAM is extensively experimentally validated on synthetic and real data. Keywords: Causal Inference, Generative Adversarial Networks, Graphical Models |
| Researcher Affiliation | Collaboration | Diviyan Kalainathan EMAIL Fentech 20 Rue Raymond Aron, Paris, France Olivier Goudet EMAIL Univ Angers, LERIA, SFR MATHSTIC, F-49000 Angers, France Isabelle Guyon EMAIL TAU, LRI, INRIA, CNRS, Universit e Paris-Saclay 660 Rue Noetzlin, Gif-Sur-Yvette, France David Lopez-Paz EMAIL Facebook AI Research 6 Rue M enars, 75002 Paris Mich ele Sebag EMAIL TAU, LRI, INRIA, CNRS, Universit e Paris-Saclay 660 Rue Noetzlin, Gif-Sur-Yvette, France |
| Pseudocode | Yes | Algorithm 1 The Structural Agnostic Modeling Algorithm |
| Open Source Code | Yes | 1. The source code is available at https://github.com/Diviyan-Kalainathan/SAM. |
| Open Datasets | Yes | SAM is extensively experimentally validated on synthetic and real data. ...The synthetic data sets involved in a first experimental analysis are DAGs with 20 or 100 variables. ...The generators Sigmoid AM, GP AM and GP Mix used for the validation of the CAM algorithm (Peters et al., 2014) can be found at https://github.com/cran/CAM. ...The SYNTREN (Van den Bulcke et al., 2006) and GENENETWEAVER (GNW) (Schaffter et al., 2011) simulators of genetic regulatory networks have been used to generate observational data... ...The largest three networks of the DREAM5 challenge (Marbach et al., 2012) are considered to assess the scalability of SAM. |
| Dataset Splits | No | The paper states: "For the sake of robustness, 16 independent runs have been launched for each data set-algorithm pair with a bootstrap ratio of 0.8 on the observational samples." This indicates a sampling strategy for evaluation runs, but it does not specify explicit training, validation, and test splits for the models themselves, nor does it provide details on how the bootstrap samples were specifically used (e.g., if a 0.8 ratio means 80% for training and the rest for validation/testing, or if it's a different use case). |
| Hardware Specification | Yes | In this work we use an Nvidia RTX 2080Ti graphics card with 12 GB memory. ...The computational training time is 113 seconds on a Nvidia RTX 2080Ti graphic card, with niter = 3000 iterations. ...CPU computational times are measured on a 48-core Intel(R) Xeon(R) CPU E5-2650 CPU. |
| Software Dependencies | Yes | The SAM algorithm is implemented in Python 3.5 with Pytorch 1.4 library for tensor calculation with Cuda 10.0. ...For AUPR and AUC evaluations, we use the scikit-learn v0.20.1 library (Pedregosa et al., 2011). |
| Experiment Setup | Yes | Each causal mechanism ˆfj is sought as a 2-hidden layer NN with 20 neurons, using tanh activation. The discriminator is a 2-hidden layer NN with n D h = 200 Leaky Re LU units on each layer and batch normalization (Ioffe and Szegedy, 2015). Structural gates ai,j are initialized to 0 with probability 1/2, except for the self-loop terms ai,i set to 0. SAM is trained for niter = 3, 000 epochs using Adam (Kingma and Ba, 2014) with initial learning rate 0.01 for the generators and 0.001 for the discriminator. In all experiments, we set the acyclicity penalization weight to λD = 0 if t < 1, 500, 0.01 (t 1, 500) otherwise. ...The hyper-parameter configuration is set to (λS = 0.02, λF = 2 10 6) in the comparative benchmark evaluation presented in next section. |