DoDiscover is a Python library for causal discovery (causal structure learning). If one does not have access to a causal graph for their modeling problem, they may use DoDiscover to learn causal structure from their data (e.g., in the form of a graph).
Why do we need another causal discovery library? Here are some design goals that differentiate DoDiscover from other causal discovery libraries.
An analyst should be able to get a causal discovery workflow working quickly without intimate knowledge of causal discovery algorithms. DoDiscover prioritizes the workflow over the algorithms and provides default arguments to algorithm parameters.
Many cutting-edge causal discovery algorithms rely on deep learning frameworks. However, deep learning-based causal discovery often requires obscure boilerplate code, complex configuration, and management of large artifacts such as embeddings. DoDiscover seeks to create abstractions that address these challenges and make deep causal discovery more broadly accessible. Current algorithms are a work-in-progress. We will begin by providing a robust API for the fundamental discovery algorithms.
Domain experts bring a large amount of domain knowledge to a problem. That domain knowledge can establish causal assumptions that can constrain causal discovery. Causal discovery (indeed, all causal inferences) requires causal assumptions.
However, a newly developed causal discovery algorithm has a greater research impact when it can do more with fewer assumptions. This "do more with less" orientation tends to deemphasize assumptions in the user interfaces of many causal discovery libraries.
DoDiscover prioritizes the interface for causal assumptions. Further, DoDiscover seeks to help the user feel confident with their assumptions by emphasizing testing assumptions, making inferences under uncertainty, and robustness to model misspecification.
DoDiscover is a Python library for causal discovery (causal structure learning). Our goal is to provide developers and researchers with guide rails for causal discovery that doesn't require deep knowledge of individual causal discovery algorithms.
The goal of dodiscover is to flatten the on-ramp to causal discovery algorithms. DoWhy provides a consistent API for various causal tasks that typically require a graph structure. Similarly, DoDiscover aims to provide a cohesive and user-friendly API to apply causal discovery algorithms for inferring a causal graph from data.
causal-learn is an extensive collection of causal discovery algorithms. It continuous to host new cutting-edge algorithms in causal discovery. However, these algorithms do not have a unified API. Further, the historic focus of causal-learn is increasing the capabilities of discovery algorithms. In contrast, dodiscover's focus is on the discovery API and usability.
When possible, dodiscover prefers to provide an API wrapper to discovery algorithms in causal-learn and other libraries. Please consider contributing to causal-learn if you plan to implement an algorithm from scratch, then contributing a wrapper in dodiscover.
In the future we plan on trying to integrate the two libraries.
pywhy-graphs is the home of graph data structures and graph algorithms in PyWhy.
py-indep serves as a repository for implementations of (un)conditional independence tests, which can be utilized in various tasks, such as causal discovery.
We are excited to welcome causal-learn to the PyWhy community. We are currently incorporating do-discover innovations into causal-learn and integrating causal-learn with other PyWhy libraries. This will occur over the course of many release cycles though. In the meantime, feel free to open up issues/PRs related to API and algorithm issues you find.
See the development version documentation.
Or see stable version documentation
Installation is best done via pip or conda. For developers, they can also install from source using pip. See installation page for full details.
Minimally, dodiscover requires:
* Python (>=3.8)
* numpy
* scipy
* networkx
* pandas
For explicit graph functionality for representing various causal graphs, such as ADMG, or CPDAGs, you will also need:
* pywhy-graphs
* graphs # this is a development version for PRable MixedEdgeGraph to networkx
For explicitly representing causal graphs, we recommend using pywhy-graphs package, but if you have a graph library that adheres to the graph protocols we require, then you can in principle use those graphs.
If you already have a working installation of numpy, scipy and networkx, the easiest way to install dodiscover is using pip:
# doesn't work until we make an official release :p
pip install -U dodiscover
To install the package from github, clone the repository and then cd into the directory. You can then use poetry to install:
poetry install
# for graph functionality
poetry install --extras graph_func
# to load datasets used in tutorials
poetry install --extras data
# if you would like an editable install of dodiscover for dev purposes
pip install -e .