LSU ECE Professor’s Paper Accepted Into Prestigious AAAI Conference

LSU Electrical & Computer Engineering Associate Professor Xiangwei Zhou recently had his paper, titled “DualGFL: Federated Learning with a Dual-Level Coalition-Auction Game,” accepted into the 2025 AAAI Conference on Artificial Intelligence, an annual event that brings together researchers, practitioners, and industry leaders to share cutting-edge advancements in AI. 

Federated learning (FL) has emerged as a powerful approach to decentralized machine learning, allowing multiple clients to collaboratively train models while preserving data privacy. Game-theoretical methods have been explored to optimize federated learning strategies. However, most existing studies rely on single-level games in either cooperative or competitive settings, failing to fully capture the complex interactions among participants.

To overcome this limitation, Zhou and his co-authors at LSU and University of Louisiana Lafayette introduced DualGFL, a novel FL framework that integrates a dual-level game in cooperative-competitive environments. This approach improves system efficiency and fairness by strategically modeling both client coalition formation and competition for training participation.

At the lower level, DualGFL employs a hedonic game, where clients form coalitions based on their preferences. The authors introduce an auction-aware utility function to guide coalition formation and implement a Pareto-optimal partitioning algorithm. This ensures that the coalition structure is optimized to balance utility among participants.

At the upper level, coalitions participate in a multi-attribute auction game to bid for training participation. This phase accounts for resource constraints and aims to derive equilibrium bids that maximize both coalitions' winning probabilities and profits. A greedy algorithm is proposed to optimize the central server’s utility, ensuring efficient resource allocation.

By integrating cooperative and competitive dynamics, DualGFL establishes a more robust federated learning paradigm. The combination of hedonic coalition formation and multi-attribute auctions ensures fair participation and efficient resource utilization, paving the way for scalable and effective decentralized learning systems.

On another note, Zhou’s PhD student Xiaobing Chen, who contributed to the paper, has officially accepted an offer to join Meta (formerly Facebook, Inc.) after four years of dedicated PhD studies at LSU.