Putting a Human Face on the Algorithm: Co-Designing Recommender Personae to Democratize News Recommender Systems

Abstract

Algorithmic recommender systems are on the rise in various societal domains, including journalism. While they offer great promise by making useful selections of large content pools, they raise various ethical and societal concerns due to their alleged lack of transparency, diversity and agency. Especially in the news context, this has serious implications because access to information is crucial in democratic societies. In this article we empirically explore the idea of algorithmic recommender personae as a productive socio-technical solution to these problems. We present the results from a two-phased qualitative study with Dutch and Belgian news readers (N ¼ 27) to 1) co-design potential news recommender personae by inductively discerning core news reading motivations and relevant features, and 2) evaluate the most promising personae on their usefulness. Results highlight three distinct recommender personae (Expert, Challenger and Unwinder) that correspond with news consumers' most salient reading motivations. We conclude that, in an increasingly automated future, allowing users more control and including them when designing recommender systems is key. With this study we hope that media organizations take up the challenge towards developing human-centered and responsible algorithmic systems that serve the public good.

Measuring the benefit of increased transparency and control in news recommendation

“…An example is Overlap Word-Cloud (see Figure 4), which explains video summaries by describing the coverage and prominence of concepts both in relation to a full-length video and within the summary itself (Inel, Tintarev, and Aroyo 2020). Another approach has been to explain user preferences by comparing and contrasting parts of the profile (Balog, Radlinski, and Arakelyan 2019). This approach uses a set-based recommendation technique that permits the user model to be explicitly presented to users in natural language.…”

“…One way to design goal‐focused exploration tools for news would be to leverage recent work that has identified three distinct recommender personae (Expert, Challenger and Unwinder) that correspond with news consumers' most salient news reading motivations (den Bogaert, Geerts, and Harambam 2022). While the Unwinder represented users aiming to relax and find entertainment; the Challengers were more keen to explore and learn different takes and angles on a topic; while Experts aim to deepen their knowledge through profound and high‐quality news articles, but not provocative or opinionated.…”

Contesting Personalized Recommender Systems: A Cross-Country Analysis of User Preferences

“…Many scholars, therefore, promote the view that users should have some control over RS (Harambam et al, 2018(Harambam et al, , 2019He et al, 2016;Jannach et al, 2017;Van den Bogaert et al, 2022). Jannach et al (2017) discuss different mechanisms that can be implemented to facilitate users to contest RS.…”

“…Three phases can be distinguished where users can contest an RS. These consist of focusing on the input, i.e., altering personal information to change how a recommendation might be personalized, the process, i.e., changing the parameters of the algorithm, or the output, i.e., modifying the order of the results (Jannach et al, 2017;Van den Bogaert et al, 2022). As such, there are multiple ways for users to potentially tweak the RS.…”

“…One promising approach are so-called algorithmic recommender personae which offer users an "intuitive, one-click option to select different sorts of personalized (news) recommendations based on their specific mood, interest and purpose" ( Van den Bogaert et al, 2022, p. 2). By selecting a certain type of recommender system, the user receives different recommendations based on a different underlying algorithmic design (Harambam et al, 2019;Van den Bogaert et al, 2022). Empirical research suggests that users would appreciate a choice between personalized and non-personalized RS (Harambam et al, 2019;Monzer et al, 2020).…”