Machine Learning for Subject Classification of Books Using Open Bibliographic Metadata

This study evaluates automatic subject classification of books using open bibliographic metadata from Open Library. A balanced dataset of 15,000 records was constructed across ten broad subject categories, using only titles and descriptions as predictive input, while original subject labels were used only to derive target classes. Classical TF-IDF-based classifiers were compared with MiniLM sentence embeddings followed by Logistic Regression. The transformer-based representation achieved the best performance, with a macro-F1 score of 0.6173, outperforming the strongest classical baseline. The results indicate that semantic embeddings can support broad subject-category prediction, although expert review remains necessary for ambiguous records.