A metadata-assisted cascading ensemble classification framework for automatic annotation of open IoT data

Public Internet of Things (IoT) platforms, such as Thingspeak, significantly increased the availability of open IoT data and enabled faster and cheaper development of novel IoT applications by reducing or even eliminating the need for deploying their own IoT sensors and platforms. However, open IoT data is often heterogeneous, sparse, fuzzy, and lacks accurate description (which we refer to as IoT metadata). These limitations make open IoT data challenging to integrate and use, and prevent the efficient development of IoT applications. In fact, while several sensor data description models have been proposed and standardized, open IoT data currently lack or include only partial metadata description. Therefore, novel techniques for automatically annotating open IoT data are needed to fully unleash the power of open IoT. This paper proposes a novel Metadata-Assisted Cascading Ensemble classification framework (MACE) for the automatic annotation of IoT data. MACE is capable of sequentially combining standalone classifiers, enabling it to cope with heterogeneous IoT data and different domains of information (e.g. numerical and textual), which have not been considered previously. MACE incorporates a novel ensemble approach for automatically selecting, sorting, filtering, and assembling classifiers in a way that improves annotation performance. The paper presents extensive experimental evaluations of MACE using public IoT datasets. Results demonstrate that the MACE framework significantly outperforms existing solutions for open IoT data by as much as 10% in classification accuracy.