MaRCA: Multi-Agent RL for Recommender Systems

This paper addresses a crucial problem in modern recommender systems: efficient computation allocation to maximize revenue. It proposes a novel multi-agent reinforcement learning framework, MaRCA, which considers inter-stage dependencies and uses CTDE for optimization. The deployment on a large e-commerce platform and the reported revenue uplift demonstrate the practical impact of the proposed approach.

• •Proposes MaRCA, a multi-agent RL framework for computation allocation in recommender systems.
• •Employs CTDE for end-to-end optimization.
• •Introduces AutoBucket TestBench and MPC-based Revenue-Cost Balancer.
• •Achieved a 16.67% revenue uplift in a real-world deployment.