Predictive Model Resilience in Edge Computing

Qiyuan Wang, Jordi Mateo-Fornés, Christos Anagnostopoulos, Kostas Kolomvatsos

Abstract: Node failure is a commonly seen threat in distributed Machine Learning systems. It is hard to predict having a huge negative impact on system availability to provide e.g., predictive analytics. Considering the benefits obtained from reduced latency and bandwidth overhead in Edge Computing (EC), invocation of the Cloud should be avoided. Hence, finding the best substitute nodes at the network edge to be invoked instead of failing nodes, evidently, builds the system's resilience upon node failures. To achieve this goal, we contribute with a resilience mechanism that relies on several data-mixing strategies that build enhanced models in each node. Such models have satisfactory prediction capabilities to handle failing nodes' predictive tasks, thus, ensuring resilience in predictive services. Furthermore, we propose a graph-driven approach to guide node invocation minimising the performance loss upon node failures. Our performance evaluation and comparative assessment showcase the applicability of our model resilience approach in intelligent EC.