M
Medibot
Text
Long-Term Fairness Analysis of ML Systems by Simulating Their Feedback Loop
FairSense's contribution: instead of checking a model for fairness at a single moment, model the deployed system and its environment as one closed loop and simulate it forward. The loop is specified as an ML model, a decision maker, a stateful environment, a stochastic distribution-shift function saying how a decision perturbs the population, and a projection function saying what the system observes next round. Monte-Carlo simulation over each configuration produces evolution traces, scored by average or maximum increase in unfairness relative to the starting state; covering-array sampling keeps the exponential configuration space tractable; and regression-plus-ANOVA sensitivity analysis with eta-squared effect sizes ranks which system parameters and which environmental uncertainties actually move long-term fairness. The point is design-time triage of where to spend effort, not prediction of what will happen; its limits are that conclusions are only as good as the assumed shift model and that it covers pairwise, not higher-order, parameter interactions.