Causal Graphical Models

Encoding causality into the CF explanation generation.

Distributions

https://github.com/amirhk/recourse/blob/master/distributions.py

---

source

BASEDISTRIBUTION

CLASS relax.data.scm.BaseDistribution (name)

Base class for all distributions.

---

source

MIXTUREOFGAUSSIANS

CLASS relax.data.scm.MixtureOfGaussians (probs, means, vars)

Mixture of Gaussians distribution.

---

source

NORMAL

CLASS relax.data.scm.Normal (mean, var)

Normal distribution.

Load SCM

Load scm structural equations.

---

source

SANITY_3_LIN

relax.data.scm.sanity_3_lin ()

Causal Model

---

source

CAUSALMODEL

CLASS relax.data.scm.CausalModel (scm_class)

Class with topological methods given a structural causal model.

Parameters:

• scm_class (str) – Name of the structural causal model.

Credit goes to the CARLA implementation.

scm = CausalModel('sanity_3_lin')

assert scm.get_children('x1') == {'x2', 'x3'}
assert scm.get_parents('x3') == ['x1', 'x2']
assert scm.get_ancestors('x3') == ['x1', 'x2']
assert scm.get_descendants('x1') == ['x2', 'x3']
assert scm.get_non_descendants('x1') == set()

Generate synthethic Data

Adapted from Carla.

(Experimental) Data Module

scm = CausalModel('sanity_3_lin')
df_endogenous, df_exogenous = _create_synthetic_data(scm, 1000)
d_config = TabularDataModuleConfigs(
    data_dir=".",
    data_name='sanity_3_lin',
    continous_cols=scm._continuous,
    discret_cols=scm._categorical,
)

dm = TabularDataModule(d_config, data=df_endogenous)

setattr(dm, 'scm', scm)
setattr(dm, 'exogenous', df_exogenous)

assert isinstance(dm.scm, CausalModel)
assert isinstance(dm.exogenous, pd.DataFrame)