Skip to main content

A package for creating causal graphs in pyro

Project description


This package is for creating Structural Causal Models (SCMs) in Pyro and evaluating various conditions with those models.

The causal model (example) can be created from a list of BEL statements strings (causal_graph.str_graph;, a pyBEL graph (causal_graph.bel_graph;, or a json file created by exporting a causal graph from Causal Fusion (causal_graph.cf_graph; Each causal model consists of nodes connected by directed edges. Each node then has parameters defining the distribution of that node's variables conditioned on the values of the parent nodes. These parameters are learned from data -- each graph can learn these parameters either using Maximum Likelihood Estimation (for point estimates) or Stochastic Variational Inference (for Bayesian estimates). Currently, Bernoulli, Normal, Lognormal, Exponential, and Gamma output distributions are supported; the choice of distribution is either specified during the initialization or defaults to a hard-coded mapping from BEL object types to distributions.

Once the training process is complete, the causal model can be queried in several different fashions. The basic query is to sample all of the nodes of the model and return a dictionary of node names and samples (example.model_sample). Using built-in Pyro functionality, the model can then leverage this to calculate conditioned samples (example.model_cond_sample), interventional samples using the do-calculus (example.model_do_sample, example.model_do_cond_sample), and counterfactual samples (example.model_counterfact).

The package includes a method to calculate the Conditional Mutual Information of a target node with respect to a test node of interest (example.cond_mut_info). This calculation relies only on the input data, not the model itself. However, the SCM also has a built-in method to perform the G-test on a variable of interest (example.g_test) to determine if the SCM sufficiently captures the distribution represented by the provided data. Note that performing both of these calculations requires binning the data to produce discrete distributions.

With the various methods for sampling conditional, interventional, and counterfactual distributions from the model, the SCM can estimate the Controlled Direct Effect (example.cd_effect), the Natural Direct Effect (example.nd_effect), and the Natural Indirect Effect (example.ni_effect). Finally, the SCM can write itself to a json file that can then be imported directly to Causal Fusion (example.write_to_cf).

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

bel2scm-0.0.4.tar.gz (11.9 kB view hashes)

Uploaded source

Built Distribution

bel2scm-0.0.4-py3-none-any.whl (23.2 kB view hashes)

Uploaded py3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page