STRUCTURE-PRESERVING PROCESS CASE EMBEDDINGS VIA DIRECTED GRAPH CONVOLUTIONAL NETWORKS

초록

This paper aims to develop a structure-aware embedding method for individual process cases by modeling them as directed graphs that preserve process flow and sequence-level variation. We propose a method for embedding individual process cases as directed graphs to preserve execution structure and behavioral semantics. Each case is modeled from raw event logs, with events as nodes and temporal transitions as edges, and encoded using a pretrained Dir-SAGE model-a direction-aware variant of Directed Graph Convolutional Networks (DGCNs). Applied to the BPI Challenge 2019 dataset, our approach outperforms text-based embeddings in unsupervised clustering, achieving a silhouette score of 0.489. Results demonstrate that structure-aware graph embeddings capture process-level variations more effectively, enabling interpretable and scalable case-level analysis. © 2026 ICIC International.

키워드