IR-for-Good Paper I

Recommender systems (RS) increasingly mediate access to crucial life spheres like employment and education, yet most remain optimized for short-term engagement rather than long-term user empowerment. While recent trustworthy RS research has addressed fairness, diversity, and transparency, these efforts remain fragmented, often targeting specific pipeline stages in isolation and lacking a unifying framework to realize their aggregated potential. This paper argues that the Capability Approach (CA) provides this necessary unifying framework, offering normative common ground to guide concrete interventions. Our contribution is twofold. First, we introduce the first formal mapping between RS components and CA constructs, addressing common critiques of the CA as being too abstract for technical operationalization. This mapping reveals key structural and recursive gaps that systematically constrain users' substantive freedoms in mainstream RS. Second, we complement this theoretical analysis with actionable design and evaluation implications, such as why capability-aware evaluation must move beyond engagement metrics and how participatory methods are essential to re-orient RS pipelines. We conclude by outlining a research agenda that acknowledges practical implementation challenges and calls for interdisciplinary interventions to pursue genuine user empowerment.

Public libraries can potentially benefit by automated recommendation to increase patron engagement. Unlike commercial platforms, these publicly funded services have a responsibility to remain alert to biases that emerge in algorithmic recommendations. In this work, we conduct a systematic evaluation of popularity and author nationality bias in Booklens, a non-personalized item-to-item recommender used in Danish public libraries that receives loans, clicks, mood tags, and creator name as input. We prompt the system with a set of 10,000 books of varying popularity and author nationality and analyze the resulting recommendations under different parameter configurations controlling (e.g., recommendation list length). For each configuration, we measure overall popularity bias and resulting bias toward authors of different nationalities. Our analysis reveals that popularity strongly drives recommendation outcomes, with up to a 40% relative decrease in exposure for less popular nationalities depending on the parameter setting. We also show that certain configurations significantly amplify or decrease these disparities. These results highlight the need for systematic bias analysis to be a structural component of recommender systems evaluation, supporting libraries in aligning algorithmic behavior with social and cultural missions.

The concept of relevance in Information Retrieval (IR) has been extensively studied. However, most mainstream IR models have been developed with adult users in mind, assuming cognitive maturity and autonomous interaction. Younger searchers, who increasingly integrate IR systems into their information-seeking practices, differ in cognitive abilities, information needs, and limited digital knowledge, which shape how they judge relevance, often diverging from traditional definitions assumed to work for adults. This calls for a deeper understanding of how this underrepresented group judges online content. In this study, we explore how children interpret and determine relevance when searching for information online in primary school classrooms. As information-seeking in this context is often guided by teachers, we also probe their criteria for relevance. By comparing both perspectives, we uncover points of alignment and divergence. These findings contribute to revisiting the concept of relevance for the primary school context and, more broadly, to the design and evaluation of equitable, context-aware IR systems that support responsible and inclusive information seeking practices.

Generative search engines are reshaping information access by replacing traditional ranked lists with synthesized answers and references. In parallel, with the growth of Web3 platforms, incentive-driven creator ecosystems have become an essential part of how projects build visibility and community by rewarding creators for contributing to shared narratives. However, the extent to which exposure in generative search engine citations is shaped by external attention markets remains uncertain. In this study, we audit the exposure for 44 Web3 projects. First, we show that the creator community around each project is persistent over time. Second, project-specific queries reveal that more popular voices systematically receive greater citation exposure than others. Third, we find that larger follower bases and projects with more concentrated creator cores are associated with higher-ranked exposure. Together, these results show that generative search engine citations exhibit exposure bias toward already prominent voices, which risks entrenching incumbents and narrowing viewpoint diversity.

Recommender systems in recruitment platforms involve two active sides, candidates and recruiters, each with distinct goals and preferences. Most existing methods address only one side of the problem, leading to potential inefficient matches. We propose a two-sided fusion framework that jointly models candidate and recruiter preferences to enhance mutual matches between recruiter and candidates. Additionally, we propose a personalized two-sided fusion approach to enhance fairness job recommendation. Experiments on the XING recruitment dataset show that the proposed approach improves fairness and compatibility, demonstrating the benefits of incorporating two-sided preferences into fairness-aware recommendation.