Fairness in ICML 2019 : “Optimal Decision Making under Strategic Behaviour”

Imagine one day, a huge number of decisions are made by machines. What would bother you?

Fairness. The systems are data-driven, but the data contain many biases. When it comes to the sensitive groups/weak groups, that group people say “Unfair!” Well, what would be the fairness in such systems, and how can we equip the system with fairness which is a wish of human rather than the truth.

In the recent two years, many works have focused not only on the current optimal benefits but one-step [xx] and long-term benefits [xx]. In a long-term manner, there are two settings for fairness by now: 1) unknown underlying distribution of the true population, sequential decisions influence the way of revealing the true distribution[xx]; however, there is not influence of decision on the true population ; 2) Decisions change the true distribution or states of people. In the second setting, the impact of decisions is described by dynamics [xx].

In this post, I introduce the problem setting and formulation of the two recent works that I guess under NeurIPS review. Later on I will introduce their experiments and discussions in the next post and technical details in the next next post. Both of them are Great! Good luck! Consequential ranking algorithms and long-term welfare and Optimal Decision Making Under Strategic Behavior.

Before diving into the papers, I take my questions/wishes: 1) the identification of the dynamics/impact of the decision 2) the discussion of the equilibrium states; 3) the evaluation of dynamic system would be interesting to see, the real dataset, the synthetic setting, and the analysis.

The problem setting is a building block of the fairness work in Machine Learning. The problem setting introduces the belief, which shows how they think the fairness in society. Based on that, the problem formulation transforms the wish into a tool language. The formulation shows the assumptions, the form of decision (stochastic/deterministic form), the distribution of population attributes (distribution of scores/skills), and the target (optimization task). The most important element in this line of works is dynamics, the impact of the decision on the population.

The first taking away message is “dynamic design”. Instead of giving the transition probability, the next step distribution should satisfy some property, such as maximize people themselves benefits. In the Optimal Decision Making under Strategic Behaviour, the dynamics are represented in the optimal transport way. The distribution of population attribute is transported from the previous distribution after making the transparent decision. Unfortunately, the setting is a one-step model, which focuses on the immediate utility. I hope there would have discussion or proposition to show the property of long-term impact/equilibrium states/ convergence. By the way, to see the orthogonal setting with unknown true distribution, here the true distribution is given with is random variable “y”.