Statistics Seminar: Active Learning for Fair and Stable Allocations

Presenter: Riddhiman Bhattacharya, Postdoc, UCSC

Description: We propose an active learning approach for dynamic fair resource allocation problems. In contrast to prior work that assumes full feedback from all agents on their allocations, we focus on scenarios where feedback is available only from a carefully select subset of agents at each epoch of the online resource allocation process. Despite this limitation, our algorithms achieve sub-linear regret in the number of time-periods for multiple fairness metrics commonly used in resource allocation problems and stability constraints inherent to matching mechanisms. The core innovation of our approach lies in the adaptive identification of the most informative feedback through dueling upper and lower confidence bounds. This strategy enables efficient decision-making with limited feedback, achieving favorable outcomes across various problem classes.

About the speaker: I am Riddhiman Bhattacharya, currently a postdoc at UCSC, Statistics Department, working with Justin (Sangwon Hyun). I have previously been a postdoc at Purdue and have obtained my PhD from the University of Minnesota in Statistics. I am interested in methodological development in statistics with varied applications including oceanography, biology and economics. I am also interested in theoretical development of statistics particularly in the fields of Markov Chain Monte Carlo, Optimization and Fast Sampling.