Opening Ottoman Miniature Collections Through Computational Access: Methodological Lessons for Digital Humanities

Ottoman miniature collections represent a layered cultural heritage that combines artistic, historical, social, and intellectual dimensions. Yet, access to their content is often limited by catalogue records that provide only minimal descriptive information. This discussion paper explores how computational access can expand the study and reuse of such collections by presenting a dataset of 435 Ottoman miniatures, enriched with detailed bibliographic metadata and expert-assigned thematic tags. An original multi-level classification scheme was developed. Tagging was carried out through a custom web interface, preparing the dataset for analysis through thematic network analysis, supervised learning, and unsupervised learning. Thematic co-occurrence analysis revealed clusters such as political figures, religious motifs, and urban settings, demonstrating the interpretive value of tagging beyond traditional cataloguing. Supervised learning with MobileNetV2 achieved 63% training accuracy and 49% validation accuracy, showing the model’s capacity to identify dominant themes while also exposing the limits of small and imbalanced datasets. Unsupervised clustering produced visually coherent groups, but low Silhouette scores and high Davies-Bouldin indices confirmed that thematic boundaries were not strongly separable through visual features alone. The findings highlight the need to combine computational methods with expert knowledge. While supervised approaches support targeted classification and unsupervised approaches enable exploratory insights, both benefit from integration with rich metadata. This study demonstrates how Ottoman miniatures can serve as a model for applying machine learning and network analysis to cultural heritage collections, offering transferable lessons for digital humanities research.