Exploiting Knowledge Graph for Multi-faceted Conceptual Modelling using GCN

Abstract

The relevant information obtained from multiple sources usually contributes to one intricate phenomenon in the industrial processes. Data fusion of different sources usually leads to more expressive and informative information than that of each single data source. Integrated information has been widely used to model a multi-faceted conceptual phenomenon, which provides a comprehensive and versatile view of understanding of the process. However, the conventional approaches concatenate feature vectors to integrate different facets, not considering the semantic gaps between them. Meanwhile, knowledge graph (KG) receives considerable attention in recent years as it comprises rich relational information among elements. Thus, KG provides a promising way to fuse multiple data sources by bridging the semantic gaps, which can be exploited in the modelling of a multi-faceted phenomenon. Inspired by the advancement of KG, we proposed an approach based on KG and a machine learning algorithm for multi-faceted modelling. Firstly, a domain-specified ontology was built to eliminate the varying distance metrics across facet boundaries, and KGs were generated by populating the data surrounding a multi-faceted phenomenon into this ontology. Secondly, the KGs were fed into a graph convolutional neural network (GCN) to learn the node features and the graph structure for graph embedding simultaneously with the shared parameters. Lastly, with the aim of multi-faceted conceptual modelling, the features obtained from the GCN model were used as inputs for machine learning algorithms to learn the hidden patterns of KGs. An experimental study using real-world data from the cold rolling process was conducted to demonstrate the feasibility of the proposed model.