Title: Recommender Systems with Fairness Considerations and Strategic Agent Dynamics.

Date: March 3, 2026.

Time: 11:30 am - 1:30pm EST

Location: Coda C0903 Ansley, 756 W Peachtree St NW

Link: https://gatech.zoom.us/j/96114293854

Krishna Acharya

Machine Learning PhD Student

School of Industrial and Systems Engineering
Georgia Institute of Technology

Thesis Committee:

Dr. Juba Ziani (advisor), School of Industrial and Systems Engineering, Georgia Institute of Technology.

Dr. Vidya Muthukumar, School of Electrical and Computer Engineering and School of Industrial and Systems Engineering, Georgia Institute of Technology.

Dr. Kai Wang, School of Computational Science and Engineering, Georgia Institute of Technology.

Dr. Jacob Abernethy, School of Computer Science, Georgia Institute of Technology.

Dr. Aaron Roth, School of Computer Science, University of Pennsylvania.

Abstract:

Recommender systems are central to how users discover content on modern digital platforms, mediating access to information, products, and media for billions of users. This dissertation studies three main themes: ensuring fairness across heterogeneous user populations, understanding strategic behavior by content producers, and the emerging paradigm of LLM-based recommendation.

We develop theoretical foundations for minimax fairness in online prediction in settings with multiple, potentially overlapping groups, introducing an oracle-efficient algorithm that reduces the groupwise fairness objective to standard online learning and provides formal guarantees across subpopulations. We then translate these insights into practical training methods for sequential recommendation via distributionally robust optimization, showing that conditional value at risk improves worst-group performance without requiring explicit group annotations.

On the item side, we model the interaction between platforms and content creators as a game, characterizing equilibrium behavior under different serving rules and showing that content specialization emerges endogenously from competition, with implications for item catalog diversity, producer utility, and user welfare.

Finally, we introduce GLoSS (Generative Language Models with Semantic Search), which combines generative query modeling with dense semantic retrieval to enable recommendation beyond direct lexical overlap, achieving state-of-the-art performance, particularly for cold-start users. We also highlight new levers for strategic manipulation in semantically aware recommenders.