LLF : A Lightweight Learning Framework for Resource‐Efficient Sports Pattern Recognition in Mobile Edge IoT

ABSTRACT
Real‐time sports pattern recognition in Mobile Edge Computing (MEC) and IoT systems demands high accuracy under stringent constraints on latency, energy and model size. Existing edge‐AI approaches often treat athletic data as generic time series, overlooking the unique kinematic structures of sports motion. To address this, we propose a Lightweight Learning Framework (LLF), which is a sports pattern aware design of modeling and system orchestration. In particular, LLF integrates (1) a multi‐scale kinematic encoder that captures periodic and burst‐like dynamics via entropy‐guided temporal pruning and kinematic energy gating, (2) sport‐adaptive neural architecture search tailored to activity granularity and hardware limits, and (3) adaptive motion‐aware sport inference with queue‐aware MEC offloading. Experiments are carried out over PAMAP2 and WISDM, which shows that LLF achieves competitive accuracy while reducing model size by up to 12 times, latency by 38% and energy by 42%. enabling efficient deployment on wearables and edge servers for real‐world sports pattern analytics.