This work contributes to a real-time, edge-centric inferential modeling and analytics methodology introducing the fundamental mechanisms for (i) predictive models update and (ii) diverse models selection in distributed computing. Our objective in edge-centric analytics is the time-optimized model caching and selective forwarding at the network edge adopting optimal stopping theory, where communication overhead is significantly reduced as only inferred knowledge and sufficient statistics are delivered instead of raw data obtaining high quality of analytics. Novel model selection algorithms are introduced to fuse the inherent models' diversity over distributed edge nodes to support inferential analytics tasks to end-users/analysts, and applications in real-time. We provide statistical learning modeling and establish the corresponding mathematical analyses of our mechanisms along with comprehensive performance and comparative assessment using real data from different domains and showing its benefits in edge computing.

这项工作为实时，以边缘为中心的推理建模和分析方法做出了贡献，引入了（i）预测模型更新和（ii）分布式计算中各种模型选择的基本机制。我们以边缘为中心的分析的目标是采用最佳停止理论在网络边缘进行时间优化的模型缓存和选择性转发，由于仅提供了推断的知识和足够的统计信息，而不是原始数据获得了高质量的分析结果，因此通信开销显着降低。引入了新颖的模型选择算法，以融合固有模型在分布式边缘节点上的多样性，以实时支持最终用户/分析师和应用程序的推理分析任务。