Fujitsu AI Ethics for Fairness

AI Ethics for Fairness is a software application that supports the detection, evaluation and mitigation of bias in AI models by analysing datasets, training models and applying fairness processing techniques. The application provides an interface that allows users to assess the fairness of datasets before developing AI models through a dataset assessment function. This step enables users to validate datasets and detect fairness issues in the input data.

Within the application, users define a target attribute and protected attributes in order to evaluate fairness metrics. Fairness can be detected for protected attributes, enabling the identification of potential bias affecting specific groups. The system also allows users to set multiple protected attributes, which makes it possible to detect intersectional bias that may be difficult to recognise.

The application includes a model training function that allows users to create machine learning models using the analysed dataset. During this step, users select the target attribute and a classification algorithm, such as logistic regression. After the model is created, the application allows users to evaluate fairness in the prediction results produced by the model.

Model evaluation includes validation of model accuracy and the use of SHAP values to analyse model behaviour. The application also provides bias mitigation through a fair processing function. In this step, users can apply mitigation algorithms, including post-processing techniques such as fair tuning.

Finally, the application allows users to compare experimental results across different models. Models can be compared using metrics such as accuracy and fairness, enabling users to select models with improved performance and fairness outcomes.