Quality-aware predictive modelling & inferential analytics at the network edge

The Internet of Things has grown by an enormous amount of devices over the later years. With the upcoming idea of the Internet of Everything the growth will be even faster. These embedded devices are connected to a central server, e.g. the Cloud. A major task is to send the generated data for further analysis and modelling to this central collection point. The devices’ network and deployed system are constrained due to energy, bandwidth, connectivity, latency, and privacy. To overcome these constraints, Edge Computing has been introduced to enable devices performing computation near the source. With the increase of embedded devices and the Internet of Things, the continuous data transmission between devices and Central Locations reached an infeasible point in which efficient communication and computational offloading are required. Edge Computing enables devices to compute lightweight algorithms locally to reduce the raw-data transmission of the network. The quality of predictive analytics tasks is of high importance as user satisfaction and decision making depend on the outcome. Therefore, this thesis investigates the ability to perform predictive analytics and model inference in Edge Devices with communication-efficient, latency-efficient, and privacy-efficient procedures by focusing on quality-aware results. The first part of the thesis focuses on reducing data transmission between the device and the central location. Two possible energy-efficient methodologies to control the data forwarding are introduced: prediction-based and time-optimised. Both data forwarding strategies aim to maintain the Central Location’s quality of analytics by introducing reconstruction policies. The second part provides a mechanism to enable edge-centric analytics towards latency-efficient network optimisation. One aspect shows the importance of locally generated analytical models in Edge Devices embracing each device’s data subspace. Furthermore, two possible ensemble-pruning methods are introduced that allow the aggregation of individual models at the Central Location towards accurate query predictions. The conclusion chapter presents the importance of privacy-efficient local learning and analytics in Edge Devices. With the aid of Federated Learning, it is possible to train analytical models for privacy-preserving data locally. Furthermore, for continuous changing environments, the parallel deployment of personalisation and generalisation for quality aware predictions is highlighted and demonstrated through experimental evaluation