Towards Fairness in Classification: Comparison of Methods to Decrease Bias

It is said that with great power comes great responsibility. Nowadays, we rely on machine learning systems to make decisions. Unfortunately these systems suffer from algorithmic biases; they often produce results that are systemically prejudiced due to erroneous assumptions in the machine learning process. Consequently these systems can contribute to increase biases in society and this is something we should avoid undoubtedly. The importance of the topic and the effect it has in the society has made it become an important research topic during the last years giving rise to different solutions. In this work, we selected three state-of-the-art techniques, decoupled classifiers, fairness constraints and adversarial learning, that claim to reduce bias in machine learning algorithms and compared their performance over different databases and fairness evaluation metrics. The obtained results show that there is no system performing the best in all aspects and databases but gives some hints to select the best option according to the objective.