Fairness and Equality of Opportunity in the Algorithmic Era

As many specific studies proved, biased data and lack of fairness evaluation criteria may typically lead algorithmic tools to discriminatory behaviour against vulnerable minorities and protected groups of our society. The transposition of philosophical concepts, like ethics and fairness, into our data-driven world is an extremely challenging operation at least for two fundamental reasons: 1. The attempts of formalization of such concepts contain the same echoes of these historic philosophical debates. 2. These concepts are relative to socio-cultural factors, and they are widely variable depending on the field of application, the country, the government, the religion, the epoch, etc. While the law is still failing in regulating the usage of algorithmic tools in the most delicate processes, data scientist community is focusing more and more on ethics, proposing different methods and different approaches. The aim is not to be universally right or create a model “more fair” than another – which is impossible to establish – but to provide a wide range of tools that well fit and easily adapt to the majority of cases. Our contribution is the successful adoption of a methodology traditionally belonging to distributive justice readapted in the algorithmic ranking system's context. This methodology focuses on the formal Equality of Opportunity (EOP) concept, also known as the nondiscrimination principle. EOP is crucial in the concept of meritocracy and its principles well fit all the areas, such as employment and education, where benefits are obtained and received thanks to the estimation of individual effort. This evaluation, however, is uneasy to accomplish due to the huge differences in the circumstances of each person, which represents the demographic and economic conditions – disparities – in which they come into the world and have no control. The method we follow is capable of estimating the degree of effort of every individual and at the same time getting rid of the different circumstances. In the modern context of a ranking system, the model we build provides a "fair" score that should guarantee the same possibility to everybody regardless of any advantageous position. We assess our model in various situations, evaluating the critical differences in terms of heterogeneity and of opportunity-loss rate, a new metric suggested by us who better highlights the results of our research.