A Short Intro to Bias in Book Recommendation

I recently completed my PhD at the Human-Centered Data Analytics group of Centrum Wiskunde & Informatica, funded by the KB [1]. During these four years, I worked on uncovering bias in book recommender systems. In this blog post, I will set the context for this research and take you through some findings. 

Recommender systems

Libraries hold vast collections that are increasingly digitized. In the KB, digitalization efforts started in the late 1990s, and currently more than 40 million pieces of reading material are present in the digital collection. This practice offers a unique opportunity to library patrons: they can remotely access the collection any time and acquire what they need, be it for informational or entertainment purposes. 

However, finding the right item is not that simple. Even if the user knows exactly what they are looking for, retrieving it successfully assumes the presence of accurate, descriptive and complete metadata and appropriate retrieval algorithms. The problem becomes harder when a user doesn’t know exactly what they want and they instead browse the collection looking for the next read that will spark their interest. This is where recommender systems are called to the rescue. Recommender systems are a category of algorithms that attempt to mathematically encode the tastes of people and the characteristics of items and match them to each other. To put it simply, they look at you, look at the world, look at you again and try to tell you what you need to read, eat, listen to or buy. 

Despite the promise of such systems, they are not without faults. The fact that they are computation-based creates an illusion of objectivity; an algorithm is seen as merely performing a series of calculations when it produces a recommendation, whereas a librarian (who is the traditional provider of reading advice) comes with their own preconceptions and blind spots. Well, this is not exactly right. Recommender systems are not simply straightforward calculations. Implementing them requires a list of human-made decisions: which algorithm and parameters to choose, what to optimize for, how many recommendations to provide, etc. Additionally, such systems attempt to detect behavioral patterns in the available data, which often means that they can only predict the past. This property makes recommender systems almost unique among other machine learning applications, say systems that try to automatically classify images into categories of animals. A photograph of a dog will forever be a photograph of a dog, whereas the fact that you read many murder mysteries last summer does not mean you’ll want to only read those forever.

With these in mind, I was hired by CWI in 2022 to do a PhD in collaboration with the KB as part of the Cultural AI Lab, which brings together researchers and practitioners from knowledge and cultural heritage institutes in the Netherlands. In this PhD, I explored the implementation of responsible recommender systems in the public library sector. Specifically, I studied the ways in which bias can creep in and propagate from the data (e.g., loan patterns in a public library) to the recommendations. 

Popularity bias

The main type of bias that has been studied in book recommendation is popularity bias. It emerges when popularity is unevenly distributed in a system, which is the most common scenario. In other words, usually most books in a library have rarely if ever been borrowed while other books have been borrowed by many users (Figure 1).

Popularity bias happens when books that are already popular are seen by the system as ‘safe’ choices and are therefore recommended even more. This creates a feedback loop. People read them because they are pushed to them, making them more popular and safer for the system.

In our work, we found that book recommendations can be very susceptible to popularity bias, for example when algorithms are trained on publicly available research data from an online book community [2]. For 3 months in 2025 I visited DBC Digital, the IT company that has developed and maintains recommender systems in the Danish public libraries. Their system is implemented in the central website of the libraries (bibliotek.dk) and is visualized in a banner which proposes similar books under every book in the collection. Through extensive auditing, we found that their recommender system puts a strong emphasis on popularity. In fact, books that have not been clicked so far have almost no way of ever being recommended [3]. 

Author nationality bias

Popularity is not a standalone property that only relates to quality, but rather a result of various properties that interact with each other. With this in mind, we also studied the relationship between popularity and author nationality. 

We found that when popular books are written by authors of a given nationality group, this group may dominate the recommendations even when the system doesn’t explicitly know the author’s nationality (or anything else about the author for that matter). In other words, popularity can act as a proxy feature of other features. In the case of the online book community, this meant that American-authored books were overrecommended by algorithms that were susceptible to popularity bias [2]. In the Danish library system, Scandinavian and Anglosphere authors took up the majority of the recommendations, while authors from more than 100 nationalities that are present in the collection were recommended disproportionately less [3].

What is next

Since March, I have been working as a Postdoctoral Researcher at CWI in collaboration with the Online Bibliotheek and NBD Biblion, this time under the guise of the project DiBiLi (Diagnosing Bias in Library recommender systems). Our goal in this project is to go one step further: not only evaluate bias in a potential recommender system trained on loan data from the Online Bibliotheek, but also to appropriately communicate the results to KB stakeholders so that they can make informed decisions regarding the development and evaluation of such systems. Stay tuned!

References

[1] Daniil, S. (2026). Bias in Book Recommendation. [PhD-Thesis - Research and graduation internal, Vrije Universiteit Amsterdam]. https://doi.org/10.5463/thesis.1746

[2] Daniil, S., Cuper, M., Liem, C., van Ossenbruggen, J., & Hollink, L. (2022). Hidden author bias in book recommendation. arXiv preprint arXiv:2209.00371.

[3] Daniil, S., Mollerup, S. H., & Hollink, L. (2026, March). Bias in Book Recommendation: A Case Study on the Danish Public Libraries. In European Conference on Information Retrieval (pp. 256-270). Cham: Springer Nature Switzerland. https://doi.org/10.1007/978-3-032-21324-2_22