The rapid spread of Machine Learning systems in contexts where a decision is required has introduced a new challenge for designers who have to interface with new considerations inherent to fairness and the possible intrinsic bias of systems. Studies show that there may be possible unwanted biases that AI systems present against people of specific groups, often underrepresented, based on race, sex, religion, or age, among other characteristics. Also when validating a model, the overall performance may not reflect those of the smaller subsets. In this thesis new data exploration techniques will be proposed for the analysis of the classification, going to search for those subgroups in which the model is underperforming.