Rewiring multi-modal knowledge graphs with GNN link prediction: insights from art history

Art history offers a complex web of relationships among artists, their works, and the cultural movements they shaped. Uncovering these connections provides valuable insights into stylistic evolution, thematic development, and creative divergence. This study presents a Unified Knowledge Graph that bridges heterogeneous data modalities—artist biographies and painting images—by integrating them into a single multi-modal representation. Semantic embeddings were derived from textual data using Large Language Models (LLMs), while visual features were extracted from artworks via Convolutional Neural Networks (CNNs). Graph Neural Networks (GNNs) were then employed to unify these embeddings within a consistent latent space, enabling link prediction and other relational tasks across modalities. Beyond its contributions to computational art history, the proposed framework offers a scalable and generalizable approach to multi-modal knowledge graph construction. It supports advanced tasks such as classification, clustering, and relationship inference across complex datasets. The methodology holds promise for broader applications in domains such as bioinformatics, social networks, and cultural analytics—where integrating diverse data sources is essential for uncovering hidden patterns and insights.