cdt 0.2.2

A Toolbox for causal graph inference

Causal Discovery Toolbox

Package for causal inference in graphs and in the pairwise settings for Python>=3.5. Tools for graph structure recovery and dependencies are included. The package is based on Numpy, Scikit-learn, Pytorch and R.

It implements lots of algorithms for graph structure recovery (including algorithms from the bnlearn, pcalg packages), mainly based out of observational data.

An example of application of the toolbox on the LUCAS dataset (on Lung cancer) using CGNNs can be found here : jupyter-notebook

Install it using pip: (See more details on installation below)

pip install cdt

Installation

The packages requires a python version >=3.5, as well as some libraries listed in requirements file. For some additional functionalities, more libraries are needed for these extra functions and options to become available. Here is a quick install guide of the package, starting off with the minimal install up to the full installation.

Note : A (mini/ana)conda framework would help installing all those packages and therefore could be recommended for non-expert users.

Miniconda installation [optional]

As miniconda is lighter and has most of the interesting functionalities of anaconda, we would recommand using Miniconda.

1. Go to https://conda.io/miniconda.html and download the lastest version of the software in 64bits. The python version here (2.7 or 3.6) does not make any importance at this stage.
2. Go to your downloaded folder and execute the installation script. For linux users, in a terminal :

$ bash Miniconda-X.sh

3. Start off by creating your first environment named 'py35': open a (new, or refresh the .bashrc) terminal and type :

$ conda create --name py35 python=3.5

4. Activate your new environment by using the command :

$ source activate py35

Note: Keep your environment activated in order to install all your libraries in your environment.

Python 3.5 or higher

If you installed a Miniconda framework at the last step, you already installed python3.5 in an environment, and therefore you can skip this step.

To check if python 3.5 is installed on your device, execute the command :

$ python --version

If not, you can install it with the command (Debian distribution):

$ sudo apt-get install python3 python3-pip python3-wheel

Install PyTorch

As some of the key algorithms in the cdt package use the PyTorch package, it is required to install it. Check out their website to install the PyTorch version suited to your hardware configuration: http://pytorch.org

Install the CausalDiscoveryToolboxPackage

The package is available on PyPi:

pip install cdt

Or you can also install it from source.

$ git clone https://github.com/Diviyan-Kalainathan/CausalDiscoveryToolbox.git  # Download the package 
$ cd CausalDiscoveryToolbox
$ pip install -r requirements.txt  # Install the requirements
$ python setup.py install develop --user

The package is then up and running ! You can run most of the algorithms in the CausalDiscoveryToolbox, you might get warnings: some additional features are not available

From now on, you can import the library using :

import cdt

Check out the package structure and more info on the package itself here.

Additional : R and R libraries

In order to have access to additional algorithms from various R packages such as bnlearn, kpcalg, pcalg, ... while using the cdt framework, it is required to install R.

Check out how to install all R dependencies in the before-install section of the travis.yml file for debian based distributions. The r-requirements file notes all the R packages used by the toolbox.

Overview

General package structure

The following figure shows how the package and its algorithms are structured

cdt package
|
|- independence
|  |- skeleton (Infering the skeleton from data, and removing spurious connections)
|  |  |- Lasso variants (Randomized Lasso[1], Glasso[2], HSICLasso[3])
|  |  |- FSGNN (CGNN variant for feature selection)
|  |  |- Network deconvolution[4]
|  |  |- Skeleton recovery using feature selection algorithms (RFECV[5], LinearSVR[6], RRelief[7], ARD[8,9], DecisionTree)
|  |- stats (pairwise methods for dependency)
|     |- Correlation (Pearson, Spearman, KendallTau)
|     |- Kernel based (NormalizedHSIC[10])
|     |- Mutual information based (MIRegression, Adjusted Mutual Information[11], Normalized mutual information[11])
|
|- generators
|  |- RandomGraphFromData (Generate a random graph similar to inputdata)
|  |- RandomGraphGenerator (Generates a random graph, can generate pairs of variables)
|  |- generate_graph_with_structure (generates a graph with a fixed structure)
|
|- causality
|  |- graph (methods for graph inference)
|  |  |- CGNN[12] method (In tensorflow, pytorch version needs revision)
|  |  |- PC[13]
|  |  |- GES[13]
|  |  |- GIES[13]
|  |  |- LiNGAM[13]
|  |  |- CAM[13]
|  |- pairwise (methods for pairwise inference)
|     |- ANM[14] (Additive Noise Model)
|     |- IGCI[15] (Information Geometric Causal Inference)
|     |- RCC[16] (Randomized Causation Coefficient)
|     |- NCC[17] (Neural Causation Coefficient)
|     |- GNN[12] (Generative Neural Network -- Part of CGNN )
|     |- Bivariate fit (Baseline method of regression)
|     |- GPI[18], PNL[19], Jarfo[20] to implement
|
|- utils
   |- Settings -> CGNN_SETTINGS, SETTINGS (hardware settings)
   |- Loss -> MMD loss [21, 22] & various other loss functions
   |- metrics -> Implements the metrics for graph scoring
   |- Formats -> for importing data formats
   |- Graph -> defines the DirectedGraph and UndirectedGraph class (see below)

Hardware and algorithm settings

The toolbox has a SETTINGS class that defines the hardware settings . Those settings are unique and their default parameters are defined in cdt/utils/Settings.

These parameters are accessible and overridable via accessing the class :

import cdt
cdt.SETTINGS

Moreover, the hardware parameters are detected and defined automatically (including number of GPUs, CPUs, available optional packages) at the import of the package using the cdt.utils.Settings.autoset_settings method, run at startup.

The graph class

The whole package revolves around using the DiGraph and Graph classes from the networkx package.

References

• [1] Wang, S., Nan, B., Rosset, S., & Zhu, J. (2011). Random lasso. The annals of applied statistics, 5(1), 468.
• [2] Friedman, J., Hastie, T., & Tibshirani, R. (2008). Sparse inverse covariance estimation with the graphical lasso. Biostatistics, 9(3), 432-441.
• [3] Yamada, M., Jitkrittum, W., Sigal, L., Xing, E. P., & Sugiyama, M. (2014). High-dimensional feature selection by feature-wise kernelized lasso. Neural computation, 26(1), 185-207.
• [4] Feizi, S., Marbach, D., Médard, M., & Kellis, M. (2013). Network deconvolution as a general method to distinguish direct dependencies in networks. Nature biotechnology, 31(8), 726-733.
• [5] Guyon, I., Weston, J., Barnhill, S., & Vapnik, V. (2002). Gene selection for cancer classification using support vector machines. Machine learning, 46(1), 389-422.
• [6] Vapnik, V., Golowich, S. E., & Smola, A. J. (1997). Support vector method for function approximation, regression estimation and signal processing. In Advances in neural information processing systems (pp. 281-287).
• [7] Kira, K., & Rendell, L. A. (1992, July). The feature selection problem: Traditional methods and a new algorithm. In Aaai (Vol. 2, pp. 129-134).
• [8] MacKay, D. J. (1992). Bayesian interpolation. Neural Computation, 4, 415–447.
• [9] Neal, R. M. (1996). Bayesian learning for neural networks. No. 118 in Lecture Notes in Statistics. New York: Springer.
• [10] Gretton, A., Bousquet, O., Smola, A., & Scholkopf, B. (2005, October). Measuring statistical dependence with Hilbert-Schmidt norms. In ALT (Vol. 16, pp. 63-78).
• [11] Vinh, N. X., Epps, J., & Bailey, J. (2010). Information theoretic measures for clusterings comparison: Variants, properties, normalization and correction for chance. Journal of Machine Learning Research, 11(Oct), 2837-2854.
• [12] Goudet, O., Kalainathan, D., Caillou, P., Lopez-Paz, D., Guyon, I., Sebag, M., ... & Tubaro, P. (2017). Learning functional causal models with generative neural networks. arXiv preprint arXiv:1709.05321.
• [13] Spirtes, P., Glymour, C., Scheines, R. (2000). Causation, Prediction, and Search. MIT press.
• [14] Hoyer, P. O., Janzing, D., Mooij, J. M., Peters, J., & Schölkopf, B. (2009). Nonlinear causal discovery with additive noise models. In Advances in neural information processing systems (pp. 689-696).
• [15] Janzing, D., Mooij, J., Zhang, K., Lemeire, J., Zscheischler, J., Daniušis, P., ... & Schölkopf, B. (2012). Information-geometric approach to inferring causal directions. Artificial Intelligence, 182, 1-31.
• [16] Lopez-Paz, D., Muandet, K., Schölkopf, B., & Tolstikhin, I. (2015, June). Towards a learning theory of cause-effect inference. In International Conference on Machine Learning (pp. 1452-1461).
• [17] Lopez-Paz, D., Nishihara, R., Chintala, S., Schölkopf, B., & Bottou, L. (2017, July). Discovering causal signals in images. In Proceedings of CVPR.
• [18] Stegle, O., Janzing, D., Zhang, K., Mooij, J. M., & Schölkopf, B. (2010). Probabilistic latent variable models for distinguishing between cause and effect. In Advances in Neural Information Processing Systems (pp. 1687-1695).
• [19] Zhang, K., & Hyvärinen, A. (2009, June). On the identifiability of the post-nonlinear causal model. In Proceedings of the twenty-fifth conference on uncertainty in artificial intelligence (pp. 647-655). AUAI Press.
• [20] Fonollosa, J. A. (2016). Conditional distribution variability measures for causality detection. arXiv preprint arXiv:1601.06680.
• [21] Gretton, A., Borgwardt, K. M., Rasch, M. J., Schölkopf, B., & Smola, A. (2012). A kernel two-sample test. Journal of Machine Learning Research, 13(Mar), 723-773.
• [22] Li, Y., Swersky, K., & Zemel, R. (2015). Generative moment matching networks. In Proceedings of the 32nd International Conference on Machine Learning (ICML-15) (pp. 1718-1727).