- Main
Learning and Socially Responsible Decision-Making with Strategic Feedback
- Chen, Yatong
- Advisor(s): Liu, Yang YL
Abstract
In recent years, the concepts of ``human-centered AI'' and ``responsible data science'' have gained prominence across multiple sectors, including academia, industry, government, and civil society. This interdisciplinary field addresses the significant challenges posed by algorithmic decision-making, particularly issues of bias, fairness, robustness, and transparency in diverse applications such as education, facial recognition, and machine translation. These challenges often originate from biased training data, inadequate inclusion of minority classes, and inappropriate algorithmic choices during pre- and post-processing steps. Moreover, all of these potential problems might be reinforced by human responses, exacerbating more severe problems in the long term as people perform multiple rounds of interaction with deployed models.
To develop machine learning systems that are truly socially beneficial, it is essential to consider human behavior, especially how individuals may strategically alter their data to influence model predictions when they understand the model's mechanics. Recognizing this, the thesis focuses on learning and decision-making within the context of strategic human feedback. Despite significant progress in related research areas, the current state of research in this field still exhibits critical gaps, including a lack of practical models of human behavior, insufficient understanding of existing algorithms' limitations in scenarios involving human responses, and overlooked potential problems in algorithmic decision-making systems.
We focus on three key objectives: understanding the social impact of decision rules, designing interventions that are both socially beneficial and sustainable and enhancing the practicality of algorithmic decision-making. By integrating theoretical and experimental methods from machine learning, algorithmic game theory, constrained optimization, and theoretical computer science, this thesis aims to demonstrate the necessity and benefits of incorporating human factors into algorithm design. Ultimately, this thesis strives to advance the field by developing algorithms, modeling frameworks, theorems, and experimental evidence that bridge the gap between theoretical research and practical implementation in responsible machine learning.