Two-step attribute reduction for AIoT networks

Abstract

The evolution of Artificial Intelligence of Things (AIoT) pushes connectivity from human-to-things and things-to-things, to AI-to-things, has resulted in more complex physical networks and logical associations. This has driven the demand for Internet of Things (IoT) devices with powerful edge data processing capabilities, leading to exponential growth in device quantity and data generation. However, conventional data preprocessing methods, such as data compression and encoding, often require edge devices to allocate computational resources for decoding. Additionally, some lossy compression methods, like JPEG, may result in the loss of important information, which has negative impact on the AI training. To address these challenges, this paper proposes a two-step attribute reduction approach, targeting devices and dimensions, to reduce the massive amount of data in the AIoT network while avoiding unnecessary utilization of edge device resources for decoding. The device-oriented and dimension-oriented attribute reductions identify important devices and dimensions, respectively, to mitigate the multimodal interference caused by the large-scale devices in the AIoT network and the curse of dimensionality associated with high-dimensional AIoT data. Numerical results and analysis show that this approach effectively eliminates redundant devices and numerous dimensions in the AIoT network while maintaining the basic data correlation.