In the following notebook we will show how you can use the CARLA library for causal recourse methods.

Causal Recourse

Data

Before we can do anything else we need some data, and more importantly a data generating process. We define this process by the Structural Causal Equations (SCM). In CARLA can be easily done by just calling the CausalModel class with the name of the SCM you want to use. Note that you can add your own SCM’s by adding it’s name to the scm_dict in carla/data/load_scm/load_scm and defining the equations in carla/data/load_scm/scm.

from IPython.display import display

import warnings
warnings.filterwarnings('ignore')

from carla.data.causal_model import CausalModel

Using TensorFlow backend.

[INFO] Using Python-MIP package version 1.12.0 [model.py <module>]
[WARNING] From /home/johan/Dropbox/Documents/Master/HiWi/CARLA/carla/recourse_methods/catalog/focus/distances.py:2: The name tf.losses.Reduction is deprecated. Please use tf.compat.v1.losses.Reduction instead.
 [deprecation_wrapper.py __getattr__]

scm = CausalModel("sanity-3-lin")
dataset = scm.generate_dataset(10000)

display(dataset.df)

	label	x1	x2	x3
0	0.0	-2.170740	1.611900	-0.000079
1	0.0	2.367498	-2.104416	-0.248596
2	0.0	1.684761	-2.919451	-1.002269
3	1.0	-2.192833	2.875651	0.526477
4	1.0	3.333677	-3.877350	-1.312688
...	...	...	...	...
9995	1.0	-1.267951	0.954710	-0.558307
9996	1.0	2.450825	-1.657112	0.300450
9997	0.0	1.721105	-1.271317	-0.649411
9998	0.0	0.881236	-1.813018	0.295015
9999	0.0	-1.704380	1.651801	-1.261734

10000 rows × 4 columns

scm.cgm.draw()

Model

Now that we have the data loaded we also need a classification model. You could define your own model, however here we will show how you can train one of our catalog models. Note that depending on your data you might need to tweak the training hyperparameters.

from carla.models.catalog import MLModelCatalog

training_params = {"lr": 0.01, "epochs": 10, "batch_size": 16, "hidden_size": [18, 9, 3]}

ml_model = MLModelCatalog(
    dataset, model_type="ann", load_online=False, backend="pytorch"
)
ml_model.train(
    learning_rate=training_params["lr"],
    epochs=training_params["epochs"],
    batch_size=training_params["batch_size"],
    hidden_size=training_params["hidden_size"],
    force_train=True
)

balance on test set 0.5141333333333333, balance on test set 0.5272
Epoch 0/9
----------
train Loss: 0.3865 Acc: 0.8229

test Loss: 0.3775 Acc: 0.8268

Epoch 1/9
----------
train Loss: 0.3736 Acc: 0.8273

test Loss: 0.3694 Acc: 0.8304

Epoch 2/9
----------
train Loss: 0.3732 Acc: 0.8297

test Loss: 0.3702 Acc: 0.8300

Epoch 3/9
----------
train Loss: 0.3720 Acc: 0.8313

test Loss: 0.3702 Acc: 0.8296

Epoch 4/9
----------
train Loss: 0.3714 Acc: 0.8295

test Loss: 0.3698 Acc: 0.8300

Epoch 5/9
----------
train Loss: 0.3715 Acc: 0.8305

test Loss: 0.3692 Acc: 0.8316

Epoch 6/9
----------
train Loss: 0.3702 Acc: 0.8341

test Loss: 0.3824 Acc: 0.8268

Epoch 7/9
----------
train Loss: 0.3708 Acc: 0.8303

test Loss: 0.3689 Acc: 0.8308

Epoch 8/9
----------
train Loss: 0.3719 Acc: 0.8305

test Loss: 0.3687 Acc: 0.8308

Epoch 9/9
----------
train Loss: 0.3701 Acc: 0.8327

test Loss: 0.3675 Acc: 0.8308

Recourse

Now that we have both the data, and a model we can start using CARLA to generate counterfactuals. In the following example we are getting negative labeled samples for which we want counterfactuals.

from carla.models.negative_instances import predict_negative_instances
from carla.recourse_methods.catalog.causal_recourse import (
    CausalRecourse,
    constraints,
    samplers,
)

# get factuals
factuals = predict_negative_instances(ml_model, dataset.df)
test_factual = factuals.iloc[:5]

hyperparams = {
    "optimization_approach": "brute_force",
    "num_samples": 10,
    "scm": scm,
    "constraint_handle": constraints.point_constraint,
    "sampler_handle": samplers.sample_true_m0,
}
cfs = CausalRecourse(ml_model, hyperparams).get_counterfactuals(test_factual)

display(cfs)

	x2	x1	x3
0	4.875700	-2.170740	-0.000079
1	-2.899361	1.684761	-1.002269
2	-7.100980	7.913797	-1.312688
3	-0.499893	0.684368	-2.310247
4	-1.516558	0.116863	0.610931