Exploring High-Throughput and Low-Energy Federated Knowledge Tracking Using Pipeline and Energy-Efficient Techniques

Federated learning, as a privacy-preserving distributed training paradigm, has been successfully applied to knowledge tracing and formed the federated knowledge tracking (FKT), which effectively facilitates industrial information development. However, the classical FKT with synchronous training faces challenges of participants’ computing capacity heterogeneity and high energy consumption. Existing works assume that participants execute FKT tasks with their computing and transmission components alternately at the default maximum operation settings, which leads to inefficient system resource utilization. Inspired by this fact, we propose a dual-task pipeline and energy-efficient FKT architecture with high throughput and low energy. First, our architecture features an efficient dual-task pipeline training strategy, in which the participants’ computing and transmission components work in parallel, thus improving the system throughput. Subsequently, we propose energy-efficient control algorithms for participants’ computing and transmission components to reduce the energy consumption. Experiments demonstrate that our approach can obtain significant prediction performance improvements under various given time and energy budgets compared to the well-known benchmarks.