Get started

This section focuses on explaining the general functionalities of the package, and how its components interact with each other. Then two tutorials are provided:

• The first one going through the main features of the package

• The second for GPU users, highlighting advanced features.

For an installation guide, please check here

Package Description

General package architecture

The Causal Discovery Toolbox is a package for causal discovery in the observational setting. Therefore support for data with interventions is not available at the moment, but is considered for later versions.

The package is structured in 5 modules:

1. Causality: cdt.causality implements algorithms for causal discovery, either in the pairwise setting or the graph setting.

2. Independence: cdt.independence includes methods to recover the dependence graph of the data.

3. Data: cdt.data provides the user with tools to generate data, and load benchmark data.

4. Utils: cdt.utils provides tools to the users for model construction, graph utilities and settings.

5. Metrics: cdt.metrics includes scoring metrics for graphs, taking as input networkx.DiGraph

All methods for computation adopt a ‘scikit-learn’ like interface, where .predict() manages to launch the algorithm on the given data to the toolbox, .fit() allows to train learning algorithms Most of the algorithms are classes, and their parameters can be customized in the .__init__() function of the class.

Note

The .predict() function is often implemented in the base class (cdt.causality.graph.GraphModel for causal graph algorithms). .predict() is often a wrapper calling sub-functions depending on the arguments fed to the functions. The sub-functions, such as .orient_directed_graph() for cdt.causality.graph models (which is called when a directed graph is fed as a second argument ), are implemented and documented in the various algorithms.

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, ran at startup.

These settings are overriddable in two ways:

1. By changing cdt.SETTINGS attributes, thus changing the SETTINGS for the whole python session.

2. By changing the parameters of the functions/classes used. When their default value in the class definition is None, the cdt.SETTINGS value is taken, by using the cdt.SETTINGS.get_default function. This allows for quick and temporary parameter change.

The graph class

The whole package revolves around using the graph classes of the networkx package. Most of the methods have the option of predicting a directed graph (networkx.DiGraph) or an undirected graph (networkx.Graph).

The networkx library might not be intuitive to use at first, but it comes with many useful tools for graphs. Here is a list of handy function, to understand the output of the toolbox’s outputs:

>>> import networkx as nx
>>> g = nx.DiGraph()  # initialize a directed graph
>>> l = list(g.nodes())  # list of nodes in the graph
>>> a = nx.adj_matrix(g).todense()  # Output the adjacency matrix of the graph
>>> e = list(g.edges())  # list of edges in the graph

Please refer to networkx ‘s documentation for more detailed information: https://networkx.github.io/documentation/stable/