Ezzat Elokda

Ezzat Elokda, Saverio Bolognani, Florian Dörfler et al.

We perform a behavioral experiment of karma, a class of mechanisms for repeated resource allocation with attractive fairness and efficiency properties, in theory. Individuals in these mechanisms bid non-tradable credits that flow from resource consumers to yielders, like karma. Human subjects recruited on Amazon MTurk are repeatedly and randomly paired to bid karma according to time-varying and stochastic individual preferences or urgency to acquire resources. Treatments varied in the dynamic urgency process (frequent moderate urgency versus sporadic high urgency) and the richness of the bidding scheme (binary versus full range). Results are benchmarked against random allocation, and karma achieves a (almost) Pareto improvement over random, despite the MTurk subjects deviating significantly from the theoretically optimal Nash bidding policy. Maximum improvement is attained by subjects that deviate from Nash by up to one karma bid unit on average, and positive improvement is attained with average deviations of up to 3-4 bid units. These findings hold across all treatments, among which no significant differences are found, with the exception of the sporadic high urgency process with binary bidding treatment being (weakly) favorable over others. These results offer behaviorally robust lower bounds for the expected performance of karma in human populations. They also provide guidance for future testing and implementation of karma mechanisms in the real world.

Nicolò Pagan, Joachim Baumann, Ezzat Elokda et al.

Prediction-based decision-making systems are becoming increasingly prevalent in various domains. Previous studies have demonstrated that such systems are vulnerable to runaway feedback loops, e.g., when police are repeatedly sent back to the same neighborhoods regardless of the actual rate of criminal activity, which exacerbate existing biases. In practice, the automated decisions have dynamic feedback effects on the system itself that can perpetuate over time, making it difficult for short-sighted design choices to control the system's evolution. While researchers started proposing longer-term solutions to prevent adverse outcomes (such as bias towards certain groups), these interventions largely depend on ad hoc modeling assumptions and a rigorous theoretical understanding of the feedback dynamics in ML-based decision-making systems is currently missing. In this paper, we use the language of dynamical systems theory, a branch of applied mathematics that deals with the analysis of the interconnection of systems with dynamic behaviors, to rigorously classify the different types of feedback loops in the ML-based decision-making pipeline. By reviewing existing scholarly work, we show that this classification covers many examples discussed in the algorithmic fairness community, thereby providing a unifying and principled framework to study feedback loops. By qualitative analysis, and through a simulation example of recommender systems, we show which specific types of ML biases are affected by each type of feedback loop. We find that the existence of feedback loops in the ML-based decision-making pipeline can perpetuate, reinforce, or even reduce ML biases.