koaning / brent Goto Github PK

View Code? Open in Web Editor NEW

27.0 4.0 5.0 1.09 MB

bayesian graphical modelling and a bit of do-calculus for discrete data.

Home Page: https://koaning.github.io/brent/

License: MIT License

Makefile 0.19% Python 23.35% Mako 0.02% Jupyter Notebook 76.44%

brent's Introduction

They're good DAGs: brent.

What it is

Brent is a small, but fun, python library that makes it easy to explore causal graphical modelling and do-calculus on systems with discrete variables. Brent is a tool that can help out when you can write a system like below, but want to write complex queries on it.

Quickstart

You can install brent via pip:

pip install brent

Next we need to have a dataset and create a graph from it. The code below demonstrates how to do this.

from brent import DAG
from brent.common import make_fake_df

dag = (DAG(dataframe=make_fake_df(7))
       .add_edge("e", "a")
       .add_edge("e", "d")
       .add_edge("a", "d")
       .add_edge("b", "d")
       .add_edge("a", "b")
       .add_edge("a", "c")
       .add_edge("b", "c")
       .add_edge("c", "f")
       .add_edge("g", "f"))
dag.plot()

Not only do we get pretty plots, but we also can build an expressive query on top of it.

from brent import Query
q = Query(dag).given(d=1).do(a=0, c=1)
q.plot()

If you're more interested in doing the inference, that's simple too.

# we can also see updated probabilities
q.infer()
q.infer(give_table=True)

Documentation

Liked the quickstart? The documentation (which is generated with pdoc3) can be found here.

Alpha Notice

NOTE! this project is in preview stages. I think I have something fun here and I've written unit tests on what I'm doing but parts are still going under review. Also there are parts of the library currently missing but which are on a roadmap:

conditional indepdence tests
api for counterfactual queries
more unit tests
datasets to start/teach with
clear logging
estimator/transformers for scikit-learn

Developing Locally

After cloning you may install brent in the virtual environment via:

$ pip install -e ".[dev]"

You can generate documentation locally by running:

$ pdoc --html --overwrite --template-dir doc-settings --http 0.0.0.0:12345 brent

brent's People

Contributors

Stargazers

Watchers

Forkers

bradib0y arose13 existeundelta maybeee18 chetanmehra

brent's Issues

implement v1 of counterfact model

this seems like a nice enough API design

q1 = Query(dag).do(a=0).given(g=1).do(b=1)
q2 = CounterFact(dag).when(q1).suppose_do(e=0).suppose_given(b=0).infer()

node names cannot have spaces

From a review from Jonas. It seems that node names cannot have spaces. Might be good to check/confirm this.

sklearn model needs to support NaN

Currently the sklearn model can be refactored a bit such that we can deal with missing values in the dataframe. Should add tests for this as well.

events that never happened

this is a graph for the RISK boardgame.

from brent import DAG, Query 
from brent.examples import generate_risk_dag

dag = generate_risk_dag()
q = Query(dag).given(losses=1).given(best_a1=1).given(best_a2=1)
q.plot()

when querying this we get an empty result

> q.infer()
{'d2': {},
 'best_d2': {},
 'd1': {},
 'best_d1': {},
 'a1': {},
 'a2': {},
 'a3': {},
 'best_a2': {},
 'best_a1': {},
 'losses': {}}

it's certainly not bad behavior but maybe a better way of dealing with this can be thought of.

q1 = Query(dag).do(a=0, c=1).given(d=1)
q1.plot()

maybe there should no longer be an arc between a and d.