Representation Learning on IoT Knowledge Graphs

In order to make the large amounts of messages generated by IoT devices in Smart Buildings interoperable, ontologies are used to represent the data as knowledge graphs (KGs). Learning over these IoT KGs can be used for various tasks, such as prediction or classification. Existing methods for KG representation learning are often evaluated on benchmark KGs and it is not explored how such methods perform on IoT KGs. The specific structure of the IoT KGs is likely to influence the quality of the representations. In this study, we investigate how the structure of IoT KGs affects the effectiveness of representation learning methods. Additionally, we look at the effect on representation quality of enriched IoT KGs, with for example temporal sequences or measurement value similarity, and the effect of the size of the IoT KGs. We perform experiments on three IoT KGs, with two representation learning methods (RDF2Vec and GCN) and two evaluation tasks (classification and value prediction). The results show that models trained with representations from enriched KGs outperform models trained with representations from original KGs on the evaluation tasks.(This article is a revised and extended version of [24]. It constitutes a significant extension with regards to the number and scale of experiments, embedding methods and evaluation tasks.)