Graph Neural Networks in PyTorch for Link Prediction in Industry 4.0 Process Graphs

Abstract

Process mining constitutes an integral part of enterprise infrastructure as its adaptability and evolution potential enhance the digital awareness of stakeholders. In the context of Industry 4.0 a mainstay of process mining is the integrity verification of process graphs. Since manufacturing typically consists of numerous operations, it follows that process mining techniques, including link prediction, must possess learning capabilities powerful enough to accurately evaluate the deviation degree from the respective template using a wide array of structural and functional attributes, including semantics in the form of labels denoting operations such as data request or human operator notification. In turn, this relies heavily on discerning higher order patterns because of the distributed nature of industrial processes. Graph neural networks (GNNs) are ideally suited for performing link prediction since they offer scalability, versatility, and geometric intuition. Two attribute sets were tested, one containing only structural patterns and one combining them with functional ones. Results with synthetic benchmark process graphs of varying complexity show that GNNs exploit the extra functional information in the form of labels to recover missing edges, themselves part of the graph structure, even when the functional attributes are noisy.