Towards fair budget-constrained machine learning
Duy Patrick Tu · Open access LMU (Ludwid Maxmilian's Universitat Munchen) · 2020
Machine learning systems are increasingly deployed in high-stake situations such as criminal justice, credit risk assessment, and medical diagnoses.With predictive decisions significantly impacting crucial aspects of individuals' life, society raised concerns about unfair treatment and discrimination by such algorithmic tools.Consequently, a growing body of research has established around fairness in machine learning.While the literature focused mainly on a setting, in which all features are ready at hand, this work focuses on a setting known as prediction-time active feature-value acquisition (AFA).Here, a decision maker can sequentially query information (features) at some cost and further makes a final prediction upon it.The aim of this cumulative thesis is to investigate algorithmic fairness in prediction-time AFA settings.The contributions of this work are twofold.First, a framework for choosing a set of confidence-based stopping criteria is proposed to redistribute information (feature) budgets among individuals.Often, individuals from underrepresented groups in the data will face a higher likelihood of erroneous decisions.Naturally, this framework encourages collecting more information for these individuals to ensure equally confident decisions.Using a calibrated probabilistic classifier, our experiments demonstrated single error parity (equal opportunity) in addition to calibration by groups.Second, staying in the AFA domain, we translate the problem into a Markov decision process and train a reinforcement learning agent to sequentially choose subsets of features that are predictive for the outcome but do not increase the demographic disparity.This is done by incorporating an adversary in the reward function, that penalizes the agent if it selects unfair features.By tuning a hyperparameter representing the magnitude of fairness, the framework is able to trade off predictive performance and fairness (demographic parity), which we confirmed experimentally.I would also like to thank Prof.