Efficient Collaborative Learning Over Unreliable D2D Network: Adaptive Cluster Head Selection and Resource Allocation

Recently, decentralized learning has been proposed for model training among mobile devices without center nodes. However, large resource overhead for model aggregation and synchronization would be incurred, which may reduce the learning performance under a given resource budget. To cope with these issues, we propose a novel cluster-based collaborative learning framework over device-to-device (D2D) network, where one device is selected as the cluster head for model aggregation. Within the proposed framework, the learning performance (evaluated by model divergence) and learning latency are analyzed with the consideration of imbalanced data and unreliable D2D communication. Then, an optimization problem is formulated to maximize the learning performance under a given latency constraint by joint cluster head selection and resource allocation. To solve this problem, a lower bound on latency constraint is first obtained for error-free model aggregation. The optimal learning performance is also derived with different degrees of data distribution. After that, an adaptive cluster head selection and resource allocation algorithm is developed for erroneous case by introducing the outage probability. Finally, comprehensive experiments are conducted on well-known models and datasets to illustrate the effectiveness of the proposed algorithm. The results show that our proposal can improve the learning performance while reducing communication and signaling overheads.