Coipee

A demo implementation of the Caipi explanatory interactive learning algorithm. Coipee implements a model which one can query to retrieve uncertain instances on which it lacks confidence, alongside an explanation of the model prediction. The user can then correct such explanation, then feed it back to Coipee to trigger an additional training guided by the explanation.

This implementation leverages feature masks as explanation, i.e., masks which can enable or disable input features.

Quickstart

Install through pip and venv:

mkvirtualenv -p python3.12 coipee

pip install coipee

Usage

Coipee revolves around a Coipee object:

barman = Coipee(
    model=base_model,  # the model to explain, e.g. a neural network
    fit_model=fit_model,  # the function to train the model, invoked after a correction
    pool=data_train,   # pool of data to measure the model's uncertainty, also used for query
    pool_labels=labels_train  # labels of the pool
)

A typical use involves querying the model for a number of uncertain instances

number_of_instances = 10
artifact = barman.query(10)
print(artifact.explanation)

and retrieve a feature mask: features important to the model are marked as True, while others as False. We can also threshold importance at different levels: the higher the threshold, the higher the required importance to mark a feature as important:

artifact = barman.query(10, threshold=0.01)
print(artifact.explanation)

Once we have our explanation, we can correct it by marking some important features as not important, and vice versa:

import copy

corrected_artifact = copy.deepcopy(artifact)

corrected_artifact.explanation[:] = False
corrected_artifact.explanation[[0, 1, 2]] = True

Here, we have simply said to the model that actually, only the features 0, 1, 2 are actually important. We can also directly retrieve differences between artifacts through the diff method:

print(f"Difference: {artifact.diff(corrected_artifact)}")

Now that we have corrected the explanation, we can feed it back to the model:

barman.stack_correction(corrected_artifact)  # adds the correction to correction stack of the model
barman.correct_model()  # triggers a training phase