Similarity-Aware ML Model Selection for Network-Device Collaboration

Abstract

The integration of artificial intelligence and machine learning (AI/ML) into 5G-Advanced and emerging 6G systems introduces significant challenges in efficiently managing model distribution between network infrastructure and user equipment (UE). In particular, conventional per-model transfer and activation can incur substantial signaling overhead and redundant on-device processing when multiple models coexist for different links, services, or devices. While current cellular standards provide baseline support for AI/ML operations, they do not explicitly address efficient management of multiple independently handled models with overlapping functional properties. This paper proposes a novel collaborative model representation framework designed to minimize signaling redundancy by identifying and leveraging structural and data-driven similarities between ML models. The framework introduces a formal relationship architecture that categorizes model similarities based on data characteristics and architectural properties, allowing a single representative model to effectively substitute for multiple target models. Extensive analysis across use cases demonstrates that the proposed framework can reduce cost of resource saving while significantly lowering on-device computational requirements. These results offer a scalable foundation for optimizing model lifecycle management and enhancing the efficiency of intelligent radio access networks.