Edge-cloud computing enables latency-sensitive and resource-efficient services by distributing workloads across heterogeneous devices and infrastructures, thereby optimizing performance and efficiency. However, the dynamic and constrained nature of edge environments makes hyperparameter optimization (HPO) for Machine Learning (ML) models particularly challenging, as it must balance predictive accuracy with training cost, energy efficiency, and stability. In this work, we propose MODE, a Multi-Objective HyperParameter Optimization (MOHPO) scheme tailored for efficient edge-based training. Our approach integrates NSGA-II with a Soft-DTW clustering to reduce redundant workload traces and accelerate convergence. At the same time, a change-point mechanism ensures robust performance across heterogeneous client data. We implement and evaluate our framework on a Kubernetes-based testbed using CODEF, our novel experimentation framework, incorporating CPU and memory traces across multiple distributions, CNIs, and deployment scenarios. Experimental results show that our approach consistently achieves comparable or superior predictive accuracy relative to generic and model-per-cluster baselines, while reducing training dataset size to $2.1-6.5\%$ and training time by up to an order of magnitude. Moreover, the stability mechanism significantly lowers performance variance across client data, enhancing robustness under dynamic conditions. These findings demonstrate that MODE enables scalable, efficient, and resilient ML training for dynamic edge environments.
Preprint version. This manuscript has been submitted to IEEE Transactions on Network and Service Management and is currently under review.
Publication Date: 2026-06-15