FAiRDAS: Fairness-Aware Ranking as Dynamic Abstract System

AI has become increasingly prominent in online matchmaking and ranking systems, where individuals are paired, ranked and recommended based on their characteristics and preferences. The need for long-term fairness in these applications has become crucial to prevent biases and discrimination. To address this, fairness-aware algorithms are commonly employed, incorporating fairness constraints into the ranking process. These algorithms use metrics and models to ensure equitable treatment across user groups. However, studying the long-term fairness properties of these approaches can be complex, posing challenges in understanding their evolution and convergence. In this study, we propose an abstract dynamic system as a solution to design and ensure long-term fairness in ranking systems. This approach provides valuable insights into system behaviour, metric interactions, and overall dynamics. By considering the ranking system as a dynamic system, we can model the evolution and interaction of fairness metrics over time. Our proposed approach enables the analysis of system properties, trade-offs, and tensions that arise when optimizing multiple fairness metrics. To validate its effectiveness, we apply this approach to real-world use case scenarios, demonstrating its practical applicability.