by Dennis Nguyen & Erik Hekman
Automated decision-making systems that offer some sort of context-dependent recommendation are virtually ubiquitous in the digital society: they help us navigate our environment (Fesenmeier et al, 2009), recommend songs and movies that we are likely to enjoy (Nair and Preethi, 2021), point to people we may (want to) know (Chamois et al, 2018), and show us products that we might not have thought of but still would be interested to purchase (Lu et al, 2015). Many of us are familiar with these ‘customer-centric’ examples, though data-based recommender systems also find widespread use in diverse professional contexts, including e.g., healthcare (Kaur et al, 2018), finance (Patel et al, 2021), product development (Wu et al, 2019), marketing (Den et al, 2020), and education (Soldatova et al, 2014). So far, artificial intelligence that learns from data to statistically estimate how likely a suggestion will meet a user’s (expressed or unexpressed) needs has proven to be remarkably versatile.
The media sector remains one of the primary domains in which data-driven recommender systems have become integral to how organisations engage with and cater to their audiences. Viewers, listeners, and readers are approached as “users” who seem to expect increasingly personalised media experiences that predict what type of content they would enjoy best. Moreover, navigating the vast expanse of available media content without the help of fast and efficient algorithms can become a very time-consuming task for many users. Unsurprisingly, developing the most accurate recommendations within a smooth, fast, and intuitive user experience is considered the key to competitive advantage in the media domain. The goal is to optimise personalisation through relevant data and clever algorithms (Falk, 2018).
This development is not limited to large streaming platforms (e.g., Netflix, Spotify, YoTube) but also concerns traditional media outlets in the private and public sectors (Michalis, 2022). For example, news websites (Karmi et al, 2018) and public service media in TV and radio broadcasting (Alvarez et al, 2020) make use of algorithms that try to provide their users with personalised content recommendations. How audiences seek and consume content changed drastically over the past decades and especially so-called “legacy media” have come under pressure to change their modus operandi for creating as well as delivering media products. Expectations towards conveniently available content-on-demand (or video-on-demand, short “VOD”) and personalisation forced many media organisations to review their top-down, real-time, and schedule-based broadcasting models (Kelly & Sorensen, 2021). This concerns news/public information, cultural output, and entertainment in equal measures.
Especially public service media (but also many news outlets with a public mission) face here a considerable ethical challenges. On the one hand, they need to remain competitive in an increasingly crowded and dynamic media landscape. Capturing and maintaining audiences attention is a difficult task. If they do not embrace novel technologies and find strategies to put them to beneficial use, they may risk becoming irrelevant for a growing part of their target audiences in the mid-to long term. On the other hand, personalisation through automated recommender systems poses a “dual dilemma”. First, there are risks for the individual user. For example, recommender systems often rely on expansive data collection that may invade personal privacy and/or can potentially manipulate users’ views on the world by excessively personalising content agendas. Second, and closely connected to this, personalisation-focused recommender systems confront public service media themselves with critical questions about their purpose and aims. Digital offerings that prioritise personalised experiences to bind users to a media service may clash with the public mission and underlying public values of an organisation, such as impartiality, diversity, inclusion, and transparency. Milano et al (2020) identify six broader ethical challenges for recommender systems that are in urgent need of solutions: inappropriate content, privacy, autonomy and personal identity, opacity, fairness, and social impact (e.g., polarisation and radicalisation). To varying degrees, all of these concern public service media and their adoption of recommender systems along the two dimensions of individual and organisational considerations.
The complexity of recommender systems as technical interventions with social, cultural, and political effects can hardly be overstated. One important question that arises is how to equip future professionals with a sufficiently flexible and diverse set of skills and expertise so that they can find a workable balance between the various benefits and risks of data-driven, algorithmic personalisation in the media sector.
Personalisation for (Public) Media in the Applied Data Science Program at Utrecht University
To tackle this challenge, Utrecht University offers a course on personalisation in the media sector within its Applied Data Science (ADS) program. Launched in 2020, ADS is an interdisciplinary study programme that allows students to specialise in a diversity of domains where data science plays an important role in shaping current and future trends. While Computer Science takes the lead, students can choose from courses in four disciplines: Health Science, Geoscience, Social and Behavioural Science, and Media Studies.
Utrecht University’s Department for Media and Culture Studies, Utrecht Data School, and the University of Applied Science’s Research Group Human Experience & Media Design joined forces to develop the course on personalisation via recommender systems in the media sector.
Following a transdisciplinary approach, the course connects different relevant disciplines to critically approach the development, design, and implementation of recommender systems for media organisations.
More specifically, the course builds on three pillars:
- Conceptual: what is the (changing) role of media in society? How are they connected to public values? What are public values in the first place and how can we define them? Why is technology never truly neutral? How can we spot and approach ethical issues? These and other critical questions are the starting point for developing recommender systems. Any technical intervention needs to fit the social, cultural, economic, and political specifics of a given context. The conceptual part is the foundation of the course that connects all the dots within the bigger picture. A guiding principle is Value-Sensitive Design, which challenges tech creators to leave silo-thinking behind and engage with a wide spectrum of stakeholders to understand how different views and needs can be reconciled. The ethical implications and societal impacts of datafication and automation are essential topics.
- Technical: recommender systems come in different forms and may utilise a variety of techniques for collecting and analysing data. There are non-personalised, content-based, and collaborative types of recommender systems. They differ in terms of the considered data points, complexity, and end-goals for recommendations (obviously, not all serve personalisation as such). Students learn how to deal with the data science part by experimenting with relevant data about media content and -consumption in Python. Decisions about specific choices in the construction of their systems are reflective of considerations on the conceptual and design dimensions.
- Design: While recommender systems mostly work hidden in the background, users encounter and engage with them through interfaces. Interaction possibilities that allow users to give some sort of (direct or indirect) input translate the resulting data into some sort of recommendation that is presented visually. How should the interfaces be designed to meet different users’ needs? How transparent should recommender systems be? What information should be offered in which ways – and should users have a greater influence on how an algorithm is “trained”? Concepts such as algorithmic affordances, the data-driven feedback loop, and user experience research are in the focus of this pillar.
The guiding question is: how can we build recommender systems that sufficiently address all stakeholders’ needs while avoiding ethical pitfalls? The dynamics involved here suggest that any design must be adaptable to constant change. Recommender systems should be built on a feedback loop that not only “learns” from users’ inputs and interaction data but also considers organisational transformations and cultural factors. The course thus places emphasis on a holistic view and stimulates students to learn more about the different social groups that are affected by their designs. Domain knowledge and sensitivity for alternative viewpoints on values, needs, risks etc. are imperative for devising fitting solutions.