Using Siamese Graph Neural Networks for Similarity-Based Retrieval in Process-Oriented Case-Based Reasoning

Abstract

Similarity-based retrieval of semantic graphs is widely used in real-world scenarios, e. g., in the domain of business workflows. To tackle the problem of complex and time-consuming graph similarity computations during retrieval, the MAC/FAC approach is used in Process-Oriented Case-Based Reasoning (POCBR), where similar graphs are extracted from a preselected set of candidate graphs. These graphs result from a similarity computation with a computationally inexpensive similarity measure. The contribution of this paper is a novel similarity measure where vector space embeddings generated by two siamese Graph Neural Networks (GNNs) are used to approximate the similarities of a precise but therefore computationally complex graph similarity measure. Our approach includes a specific encoding scheme for semantic graphs that enables their usage in neural networks. The evaluation examines the quality and performance of these models in preselecting retrieval candidates and in approximating the ground-truth similarities of the graph similarity measure for two workflow domains. The results show great potential of the approach for being used in a MAC/FAC scenario, either as a preselection model or as an approximation of the graph similarity measure.