Recommender systems are one of the most widely used applications of artificial intelligence.Their use can have far-reaching consequences for stakeholders, users, and society at large.In this third edition of the NORMalize workshop, we once again seek to advance the research agenda of normative thinking, considering the norms and values that underpin recommender systems, as well as to introduce the concept to a broader audience.We aim to bring together a growing community of researchers and practitioners across disciplines who want to think about the norms and values that should be considered in the design and evaluation of recommender systems, and to further educate them on how to reflect on, prioritise, and operationalise such norms and values.NORMalize 2025 is a half-day workshop focusing on discussion and interdisciplinary collaboration, building upon its two successful runs at previous RecSys conferences in 2023 and 2024.

Books occupy a significant cultural role in human societies, and libraries have long positioned themselves as institutions committed to equitable access to information and promotion of reading. As libraries increasingly adopt AI-driven tools, it becomes important to examine the potential for bias introduced by these systems. This thesis investigates bias in book recommender systems, with an emphasis on the public library sector. We approach this topic from three perspectives: understanding the current ethical considerations around book recommendation, examining the challenges of measuring statistical bias, and studying the societal implications of statistical bias in book recommendation. In the first part of the thesis, we map the existing landscape. We present the first systematic literature review of bias in book recommender systems, surveying 40 papers from computer science venues. We find that existing research is fragmented, often treats books as interchangeable with other media items, and rarely engages with the unique characteristics of the book domain. We propose future research directions with an emphasis on interdisciplinarity and domain specificity. Additionally, we conduct an interview study in which we explore how practitioners at three public service media organizations in the Netherlands conceptualize diversity in recommender systems. We find that diversity is subject to a wide range of interpretations even within a narrow domain, and that normative choices are unavoidable in operationalizing it. In the second part, we take a critical look at the measurement of popularity bias. We reproduce three prominent studies on popularity bias in media recommendation across the movie, music, and book domains, and identify four aspects as potential sources of divergence in results: data, algorithms, division of users in groups, and evaluation strategy. We find that all aspects contribute to the divergence, with the evaluation strategy playing a particularly significant role. We further experiment with synthetic and real data to examine the joint effect of data characteristics and algorithm configuration, finding that the presence and magnitude of popularity bias are highly sensitive to these choices. These findings indicate that conclusions on bias should be explicitly scoped to the limits of the experimentation. In the third part, we study the relationship between statistical bias and social bias. Using the Book-Crossing dataset, we show that popularity bias in collaborative filtering leads to the over-recommendation of books by American authors, whose works dominate the data. We then conduct the first audit-type study of bias in a library recommender system in production: Booklens, the non-personalized item-to-item system used by the Danish public libraries. We find that Booklens is strongly prone to popularity bias, and that this propagates into author nationality bias, with books by less popular nationalities being systematically underrepresented. We show that, while tweaking system parameters can reduce the effect, bias persists across configurations. The findings of this thesis highlight that bias in book recommendation is an underexplored yet consequential area of research. We argue that addressing it requires specificity in research design and a willingness to closely engage with the domain, and the institutions that use these systems.

The extent to which popularity bias is propagated by media recommender systems is a current topic within the community, as is the uneven propagation among users with varying interests for niche items. Recent work focused on exactly this topic, with movies being the domain of interest. Later on, two different research teams reproduced the methodology in the domains of music and books, respectively. The results across the different domains diverge. In this paper, we reproduce the three studies and identify four aspects that are relevant in investigating the differences in results: data, algorithms, division of users in groups and evaluation strategy. We run a set of experiments in which we measure general popularity bias propagation and unfair treatment of certain users with various combinations of these aspects. We conclude that all aspects account to some degree for the divergence in results, and should be carefully considered in future studies. Further, we find that the divergence in findings can be in large part attributed to the choice of evaluation strategy.