If you judge a book by its cover

University libraries hold vast collections of scholarly work, yet most academic books are borrowed only a handful of times each year. A study in the International Journal of Information and Communication Technology suggests that the problem lies less in library logistics than in the lack of a sophisticated recommendation system available to readers. The team behind the research have come up with a new approach to library recommendation systems that replaces the static models with an approach that adapts to the readers’ changing learning needs.

For decades, most library and commercial platforms have relied on collaborative filtering, a technique that recommends items based on aggregated past behaviour, such as borrowing or purchasing patterns. While effective at scale, the method treats readers as having a fixed profile. It ignores the level of difficulty of material relative to a reader’s ability. Moreover, it does not work well with a cold-start, where little data exist for new users or new books. This latest research suggests that overcoming such limitations could open up knowledge to more readers and stop those books gathering dust on the library shelves.

The new system models readers as learners whose knowledge changes over time. It uses a gated recurrent unit, a form of neural network designed for time-series data. This tracks changes in a reader’s mastery of a subject and so can produce what the researchers refer to as a continuously updated “cognitive state matrix”. This analysis reflects what a reader is likely to understand at any given moment in their education of research.

The team adds that their model incorporates behavioural signals, such as borrowing rhythms and search intent, and an environmental feedback mechanism that adjusts recommendations to balance a resource’s difficulty against its popularity.

The approach was tested using real borrowing data from a university library. The team found improvements over established baselines in ranking quality and measured learning gains, while maintaining low response times compatible with live deployment.

Deng, F. (2025) ‘Personalised book recommendation model for university libraries based on multi-factor knowledge tracking’, Int. J. Information and Communication Technology, Vol. 26, No. 50, pp.1–16.