Extension of Knowledge Graphs to Support the Visualization of Relationships

Conventional knowledge graph use is limited by explicit links, hindering the serendipitous discoveries crucial for creative innovation. Our study constructs a knowledge expansion framework that structurally induces serendipity as a reproducible process by extending latent relationships through a proposed three-stage model. As a theoretical foundation, we introduce the concepts of serendipity, abduction, and hypothetical constructive reasoning. Our approach utilizes H2GCN for link prediction and gSpan for structural pattern mining to extract novel candidate relationships. In evaluation experiments, the H2GCN Clustering condition demonstrated the most well-balanced performance regarding structural novelty, semantic diversity, and output stability, indicating its potential as a robust foundation for creative discovery support. While other methods generated more patterns, they often suffered from high variance, revealing a critical trade-off between structural coverage and stability. Key challenges identified include this trade-off and the lack of semantic interpretation for extracted patterns. Based on these insights, we plan to develop a multi-agent architecture that integrates GNN-based structural prediction with LLM-driven semantic hypothesis generation to realize a holistic, collaborative creativity support system between humans and AI.