Resilient edge predictive analytics by enhancing local models

Abstract

In distributed computing environments, the collaboration of nodes for predictive analytics at the network edge plays a crucial role in supporting real-time services. When a node’s service becomes unavailable for various reasons (e.g., service updates, node maintenance, or even node failure), the rest of the available nodes connot efficiently replace its service due to different data and predictive models (e.g., machine learning [ML] models). To address this, we propose decision-making strategies rooted in the statistical signatures of nodes’ data. Specifically, these signatures refer to the unique patterns and behaviors within each node’s data that can be leveraged to predict the suitability of potential surrogate nodes. Recognizing and acting on these statistical nuances ensures a more targeted and efficient response to node failures. Such strategies aim to identify surrogate nodes capable of substituting for failing nodes’ services by building enhanced predictive models. Our resilient framework helps to guide the task requests from failing nodes to the most appropriate surrogate nodes. In this case, the surrogate nodes can use their enhanced models, which can produce equivalent and satisfactory results for the requested tasks. We provide experimental evaluations and comparative assessments with baseline approaches over real datasets. Our results showcase the capability of our framework to maintain the overall performance of predictive analytics under nodes’ failures in edge computing environments.

Keywords

• edge computing
• model resilience
• predictive service
• model generalization
• distributed machine learning
• node failures