Ricerc@Sapienza

Algorithmic methods are the driving force behind the mechanisms of interaction between people and online markets. Examples are the hiring of workforce on the Internet, algorithmic pricing of goods, matching customers to service providers in the sharing economy, or recommending contents and products. While these tasks have a deep impact in the everyday life of people, it is clear that some of these decisions can be biased by the algorithmic methods adopted or by the data used for training the machine learning algorithms, and this is one of the main reason behind the recent daily debate on their use.

The main research problem we face is the conflict between the algorithm-designer objective of optimizing the economic benefits and the importance of providing solutions that ensure fairness. Introducing fairness as a constraint opens a whole new set of challenging problems and requires the development of new algorithmic solutions. Fairness can make algorithmic solutions for the problems considerably harder to compute and the optimal fair solution can achieve a worse objective cost than the optimal unconstrained solution.

Our approach will be the one of using algorithmic approximation and min-regret analysis of online learning methods in order to ensure that the quality of the fair solution produced by the algorithms and by the market mechanisms will be provably close to the quality of the optimal unconstrained solution when possible. We'll also investigate methods that remove algorithmic discrimination by design through the execution of a preprocessing phase before the optimization algorithm is applied. We'll focus our investigation on the problems of designing fair algorithms and mechanisms for Internet advertisement, recommender systems, online labor marketplaces, fair clustering and classification.