1. FixOut Core

Warning

FixOut is a tool centered on sensitive features and their proxies. Not providing this information correctly beforehand will compromise the obtained results.

1.1. Data distribution

FixOut shows the data distribution centered on sensitive features. Plots show the data frequency based on each possible value of a sensitive feature and the target variable \(y\).

Intersectionality. It also combines sensitive features to help users to verify the distribution take into account two sensitive features at the same time.

../_images/datadistr.PNG

Note

In the current version of FixOut, this functionality is only available using the web dashboard.

1.2. Correlation analysis

Correlation matrices are automatically computed using two coefficients: Pearson and Spearman. These matrices are computed taking into account all features: sensitive or not.

FixOut also delivers a list of non-sensitive features that are correlated to sensitive features. One list is generated for one sensitive feature in decreasing order based on the absolute value of the correlation.

Warning

Non-sensitive features that are highly correlated to sensitive features must be treated as sensitive features as well.

../_images/corr.PNG

Note

In the current version of FixOut, this functionality is only available using the web dashboard.

1.3. Fairness assessment

FixOut automatically calculates several fairness metrics given the output of a model [EU2023]. The list of all fairness metrics currently available is shown below.

  • Equalized Odds (EOD)

  • Demographic Parity (DP)

  • Equal Opportunity (EO)

  • Predictive Equality (PE)

  • Predictive Parity (PP)

  • Conditional Accuracy Equality (CEA)

See also

1.4. Reverse engineering

FixOut trains (and evaluates) various models to predict sensitive features using the target variable \(y\). This type of experiment is useful for analyzing biases. The idea is to verify whether it is possible to reconstruct sensitive features with good performance using only \(y\). See Models compatible with FixOut to check out which models can be used in this analysis.

See also

1.5. Discriminatory models

FixOut trains models using only sensitive features to check if its possible to achieve high performance using sensitive features only and then to help practionners to assess (and avoid) biases. See Models compatible with FixOut to check out which models can be used in this analysis.

See also

1.6. Models compatible with FixOut

The types of models that are compatible with FixOut are briefly shown below. FixOut mainly supports trained models for binary classification. They must be compatible with the API definition of scikit-learn.

This is not an exhaustive list, but some examples include: Linear Regression, Decision Trees, Support Vector Machines, Gaussian Naive Bayes, Random Forests, Multi-layer Perceptron, and Gradient Boosting.

Warning

FixOut is still in development. New types of models will be supported soon.

References

[EU2023]

“Survey on fairness notions and related tensions” . Alves, G., Bernier, F., Couceiro, M., Makhlouf, K., Palamidessi, C., & Zhioua, S. - 2023.