PREDICTING AND DECIDING WITH MACHINE LEARNING: (TRY TO) UNDERSTAND WHAT YOU DO AND (TRY TO) BE FAIR

Machine Learning methods to predict and decide tend to be black boxes, making it difficult to explain the output and to avoid discrimination caused by bias in the data.
Expressing the training process as a Mathematical Optimization problem allows us to control critical issues such as the amount of information used by the model, measuring the relevance of the different input data, and strengthening the robustness or fairness of the procedures and decisions. Furthermore, once the prediction model is built, Mathematical Optimization can also be used to build counterfactual solutions, by identifying how the input should be to obtain a more desirable output. This paradigm will be illustrated in this talk in different Machine Learning domains.