CERTA

Code for CERTA (Computing ER explanations with TriAngles), an algorithm for computing saliency and counterfactual explanations for Entity Resolution models.

Installation

To install CERTA locally run :

pip install .

Usage

Wrap the model whose predictions need to be explained using the ERModel interface. The get_model utility method will load an existing model, if available, or train a new one using the data in the provided dataset. E.g. for a DeepMatcher model use:

from certa.models.utils import get_model

model = get_model('dm', '/path/where/to/save', '/path/to/dataset', 'modelname')

Define a prediction function wrapping the model.predict() method.

def predict_fn(x, **kwargs):
    return model.predict(x, **kwargs)

Create a CertaExplainer. CERTA needs access to the data sources lsource and rsource.

import pandas as pd
from certa.explain import CertaExplainer

lsource = pd.read_csv('/path/to/dataset/tableA.csv')
rsource = pd.read_csv('/path/to/dataset/tableB.csv')
certa_explainer = CertaExplainer(lsource, rsource)

To generate the prediction for the first two records in the data sources, do the following:

import numpy as np
from certa.local_explain import get_original_prediction

l_tuple = lsource.iloc[0]
r_tuple = rsource.iloc[0]
prediction = get_original_prediction(l_tuple, r_tuple, predict_fn)
class_to_explain = np.argmax(prediction)

To explain the prediction using CERTA :

saliency, summary, cfs, triangles, lattices = certa_explainer.explain(l_tuple, r_tuple, predict_fn)

CERTA returns:

• the saliency explanation within the saliency pd.DataFrame
• a summary containing the set of attributes that has the highest probability of sufficiency of flipping the original prediction
• the generated counterfactual explanations within the cfs pd.DataFrame
• the list of open triangles (in form of tuples of record ids) used to generate the explanations

Examples

Examples of using CERTA can be found in the following notebooks:

• Explain DeepMatcher predictions
• Explain Ditto predictions

Citing CERTA

If you extend or use this work, please cite the paper:

@article{teofili2022effective,
  title={Effective Explanations for Entity Resolution Models},
  author={Teofili, Tommaso and Firmani, Donatella and Koudas, Nick and Martello, Vincenzo and Merialdo, Paolo and Srivastava, Divesh},
  journal={arXiv preprint arXiv:2203.12978},
  year={2022}
}