Xender$$^+$$: Enhancing Time Series Data Transmission via Historical Model Reuse

The rapid proliferation of the Internet of Things (IoT) device has led to an explosive growth of time-series data. Efficient transmission of this data to the cloud is critical but often constrained by limited network bandwidth. Existing frameworks, such as Xender, employ a sample-then-generate architecture to reduce traffic but incur high computational overhead due to frequent online model retraining. In this paper, we propose Xender \(^+\) , a novel framework that enhances transmission efficiency through historical model reuse. Xender \(^+\) maintains a cloud-side multi-instance library of pre-trained TDGen models, clustered by data patterns. Instead of retraining, incoming sampled segments are matched to existing models for direct reuse, enabling high-fidelity reconstruction. To ensure scalability, the system incorporates a dynamic cluster expansion mechanism and an LRU eviction policy for efficient library management. Additionally, a GPU-accelerated sampling algorithm at the edge significantly reduces processing latency. Extensive experiments on real-world datasets demonstrate that Xender \(^+\) consistently outperforms representative state-of-the-art baselines, including Xender and ZipMate, achieving substantially lower end-to-end latency while preserving high reconstruction accuracy.