Laura Hollink

Automated book recommendation is a compelling application given the rich and long-standing human tradition of writing and reading. The documented discrimination in the publishing world raises questions on the types of bias that can arise when algorithms are trained on user-book interactions. Various papers have studied recommender systems bias based on datasets from platforms where users share their opinions on books. In this paper, we conduct the first survey of computer science research on bias in book recommender systems. We survey 40 papers from top conferences and journals using the structured PRISMA methodology and provide an overview of the field. We map the existing literature into bias types, subjects and actions, and present key metrics and bias mitigation methods. We find that there is limited targeted research, fragmented evaluation practices, and a lack of engagement with the unique characteristics of the literary domain. Finally, we propose avenues for future work and make concrete suggestions for interdisciplinary research.

This paper examines how Large Language Models (LLMs) reproduce societal norms, particularly heterocisnormativity, and how these norms translate into measurable biases in their text generations. We investigate whether explicit information about a subject's gender or sexuality influences LLM responses across three subject categories: queer-marked, non-queer-marked, and the normalized "unmarked" category. Representational imbalances are operationalized as measurable differences in English sentence completions across four dimensions: sentiment, regard, toxicity, and prediction diversity. Our findings show that Masked Language Models (MLMs) produce the least favorable sentiment, higher toxicity, and more negative regard for queer-marked subjects. Autoregressive Language Models (ARLMs) partially mitigate these patterns, while closed-access ARLMs tend to produce more harmful outputs for unmarked subjects. Results suggest that LLMs reproduce normative social assumptions, though the form and degree of bias depend strongly on specific model characteristics, which may redistribute, but not eliminate, representational harms.

Statements on the propagation of bias by recommender systems are often hard to verify or falsify. Research on bias tends to draw from a small pool of publicly available datasets and is therefore bound by their specific properties. Additionally, implementation choices are often not explicitly described or motivated in research, while they may have an effect on bias propagation. In this paper, we explore the challenges of measuring and reporting popularity bias. We showcase the impact of data properties and algorithm configurations on popularity bias by combining real and synthetic data with well known recommender systems frameworks. First, we identify data characteristics that might impact popularity bias, and explore their presence in a set of available online datasets. Accordingly, we generate various datasets that combine these characteristics. Second, we locate algorithm configurations that vary across implementations in literature. We evaluate popularity bias for a number of datasets, three real and five synthetic, and configurations, and offer insights on their joint effect. We find that, depending on the data characteristics, various configurations of the algorithms examined can lead to different conclusions regarding the propagation of popularity bias. These results motivate the need for explicitly addressing algorithmic configuration and data properties when reporting and interpreting bias in recommender systems.

Diversity is a commonly known principle in the design of recommender systems, but also ambiguous in its conceptualization. Through semi-structured interviews we explore how practitioners at three different public service media organizations in the Netherlands conceptualize diversity within the scope of their recommender systems. We provide an overview of the goals that they have with diversity in their systems, which aspects are relevant, and how recommendations should be diversified. We show that even within this limited domain, conceptualization of diversity greatly varies, and argue that it is unlikely that a standardized conceptualization will be achieved. Instead, we should focus on effective communication of what diversity in this particular system means, thus allowing for operationalizations of diversity that are capable of expressing the nuances and requirements of that particular domain.

Web resources in linked open data (LOD) are comprehensible to humans through literal textual values attached to them, such as labels, notes, or comments. Word choices in literals may not always be neutral. When culturally stereotyping terminology is used in literals, they may appear as offensive to users in interfaces and propagate stereotypes to algorithms trained on them. We study how frequently and in which literals contentious terms about people and cultures occur in LOD and whether there are attempts to mark the usage of such terms. For our analysis, we reuse English and Dutch terms from a knowledge graph that provides opinions of experts from the cultural heritage domain about terms' contentiousness. We inspect occurrences of these terms in four widely used datasets: Wikidata, The Getty Art & Architecture Thesaurus, Princeton WordNet, and Open Dutch WordNet. Some terms are ambiguous and contentious only in particular senses. Applying word sense disambiguation, we generate a set of literals relevant to our analysis. We found that contentious terms frequently appear in descriptive and labelling literals, such as preferred labels that are usually displayed in interfaces and used for indexing. In some cases, LOD contributors mark contentious terms with words and phrases in literals (implicit markers) or properties linked to resources (explicit markers). However, such marking is rare and non-consistent in all datasets. Our quantitative and qualitative insights could be helpful in developing more systematic approaches to address the propagation of stereotypes via LOD.

AI technologies offer significant opportunities for the media sector, including content production, distribution, and audience engagement. This case study aims to bridge the technological expertise of the AI, Media & Democracy lab in the Netherlands to stakeholders in the media sector, identifying current technology implementation practices and outlining the potential of our technological expertise. First, a visual portfolio of the expertise of the researchers was created. Then, focus group workshops were held with broadcasters and media institutions, using an interactive online Miro environment. Results include insights on current implementation of the presented technologies by the media organizations, and opportunities for further implementation. Key takeaways are the broadcaster need for short-term implementation possibilities and the importance of having both B2C (broadcasters) and a B2B (the entire sector) perspectives present in focus groups to provide confirmation of findings and offer a broader viewpoint on the media landscape.

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.