Xender+ Cuts IoT Time Series Traffic by Reusing Historical Models Instead of Retraining

The rapid growth of Internet of Things devices has produced an explosion of time-series data, and transmitting that data to the cloud is often constrained by limited network bandwidth. A framework called Xender+, proposed by Bin Liu, Xuanfeng Liu, Ziguang Fu and Jian Liu in Neural Processing Letters, addresses this bottleneck by reusing historical models rather than retraining them online. The work targets a weakness in existing sample-then-generate systems such as Xender, which reduce traffic by sending only sampled segments of a data stream and reconstructing the rest, but incur high computational overhead because of frequent online model retraining.

Xender+ maintains a cloud-side multi-instance library of pre-trained TDGen models, clustered by the data patterns they capture. When a sampled segment arrives from an edge device, the system does not start a fresh training run; instead, it matches the segment to an existing model in the library and reuses that model directly, enabling high-fidelity reconstruction of the full time series. This substitution of retrieval for retraining is what allows the framework to cut end-to-end latency while preserving reconstruction accuracy, according to the authors’ experiments on real-world datasets.

Scalability is handled through two complementary mechanisms. A dynamic cluster expansion mechanism allows the model library to grow as new data patterns appear that no existing cluster represents well, so unfamiliar streams are not forced into ill-fitting models. An LRU eviction policy keeps the library manageable by removing the least recently used models, supporting efficient library management as the diversity of incoming IoT data evolves.

On the edge side, the framework incorporates a GPU-accelerated sampling algorithm that significantly reduces processing latency before data leaves the device. Because sampling is the first step in the sample-then-generate pipeline, accelerating it shortens the time an IoT device spends preparing data for transmission, so the framework targets the full end-to-end path rather than only one segment of it.

The authors report extensive experiments on real-world datasets in which Xender+ consistently outperformed representative state-of-the-art baselines, including Xender and ZipMate, achieving substantially lower end-to-end latency while preserving high reconstruction accuracy. The comparison with Xender is direct, since Xender is the prior sample-then-generate framework whose retraining overhead Xender+ is designed to avoid. The results indicate that for bandwidth-constrained IoT deployments, the cost of repeatedly training generative models online can be reduced without sacrificing the quality of reconstructed data.

The open-access paper, published on 30 October 2026, details the full architecture and experimental evaluation for researchers and practitioners working on efficient data transport for the Internet of Things.

Subject of Research: A framework for efficient IoT time series data transmission that reuses pre-trained generative models instead of retraining them online.

Article Title: Xender\(^+\): Enhancing Time Series Data Transmission via Historical Model Reuse

Article References: Liu, B., Liu, X., Fu, Z., & Liu, J. (2026). Xender$$^+$$: Enhancing Time Series Data Transmission via Historical Model Reuse. Neural Processing Letters. https://doi.org/10.1007/s11063-026-11889-4

Image Credits: AI Generated

DOI: 10.1007/s11063-026-11889-4

Keywords: Internet of Things, time series data, data transmission, model reuse, edge computing, cloud computing, bandwidth efficiency, GPU acceleration, LRU eviction, reconstruction accuracy, latency reduction, TDGen