Explainable and Energy-Efficient Selective Ensemble Learning in Mobile Edge Computing Systems

Explainable ensemble learning combines explainable artificial intelligence (XAI) and ensemble learning (EL) to solve the closed-box problem of EL and provide a clear and transparent explanation of the decision-making process in the model. As a distributed machine learning architecture, EL deploys base learners trained with local data at edge node and infers on target tasks, then combines the inference results of the participating base learners. However, selecting all base learners into EL may result in wasting more computing resources and not obtain better performance. To address this issue, we put forward the definition of confidence level (ConfLevel) on the basis of XAI and verify its effectiveness as the metric of selecting the base learner. Then, we take the joint optimization model of considering high ConfLevel and low computing power to determine the participating base learners for selective ensemble learning (SEL). Due to the non-convex and combinatorial nature of the problem, we propose a node selection and power control algorithm on the premise of Benders' Decomposition (referred to BD-NSPC) to obtain the global optimal solution efficiently. In addition, simulation results show that BD-NSPC consumes about 30% less energy per EN on average and improves accuracy by 1-2% compared to other SEL algorithms. Besides, compared with federated learning (FL) framework, BD-NSPC reduces the energy consumption by about 25% and the latency by about 28%, achieving comparable accuracy in the edge computing system.