Interdisciplinary Interventions in the Design of Music Recommender Systems
Georgina Born, Professor of Anthropology and Music, Dept of Anthropology and Institute of Advanced Studies, UCL
Fernando Diaz, Research Scientist, Google; Adjunct Professor, McGill University
Andres Ferraro, Postdoctoral Fellow, McGill University and Pandora
Gustavo Ferreira, Postdoctoral Fellow, McGill University
Music recommender systems (RS) are central to global streaming services like Spotify, Apple Music and Amazon; for many consumers they are the dominant means by which music is curated and discovered. By employing data on consumer behavior and repeatedly influencing consumer choices, RS can shape cultural literacies as well as population-wide trends in consumption and taste – yet this is largely unacknowledged in current literatures. Through the collection and mining of consumer data, RS introduce a type of ‘monitoring-based marketing’ (Andrejevic 2012) which takes place at a much larger scale and is more rapid, recursive and intensive in comparison to earlier, non-computational methods of market research. RS therefore represent a new stage in the algorithmic automation of the infrastructure for everyday music consumption (Bowker and Star 1999, Gillespie 2014).
Currently, these systems optimize for listener engagement and retention (Chen et al 2019) following a logic of ‘similarity’ enacted by ML. While this logic may increase revenues, it ignores the social and cultural implications of system design (Mehrotra et al 2018). Issues like algorithmic bias – which listeners and creators are optimized for and which are not (Baym 2013, Morris 2015, Crawford 2016a, Baeza-Yates 2018) – are symptomatic of major problems in design (Danks & London 2017, Springer et al 2018), due in part to the practice of rendering social dimensions externalities (Overdorf et al 2018, Callon 1998). Consumer bias may be introduced by the amplification of dominant groups, with likely long-term effects in shaping musical subjectivities and taste formations. For musicians, algorithmic distribution brings incentives about what type of music to produce, potentially privileging the generic over the different, auguring long-term effects on creativity (Born, Diaz et al 2021, Anderson 2014). Some of these problems are intrinsically linked to twin features of the existing recommender system paradigm – commercialisation and personalisation – as well as the model of the human (musical) subject they presuppose (Prey 2018, Stark 2018, Seaver 2022).
In this light, and to tackle these problems by pursuing their design implications, this study hosts sustained interdisciplinary dialogues between engineers designing music recommender systems and MusAI social scientists and humanists. Based on early work, four core, related elements define our research. We propose, first, that it is time for recommender design to move beyond a solely commercial orientation in favour of design paradigms oriented to human musical flourishing and the public good (Moe 2008, Andrejevic 2013, Hesmondhalgh 2013, Born 2018). Second, we propose that as well as a focus on personalisation, recommender design should acknowledge the aggregate and cumulative influence of RS and develop ways of analysing and modifying these effects – which amount to the mediation of cultural change – in progressive ways that seek to achieve the above goals. Third, as a concrete means of addressing these two issues, we have developed a new metric called ‘commonality’ which measures the degree to which recommendations familiarize a given user population with specified categories of content (Ferraro, Ferreira, Diaz & Born 2022). And fourth, as a key case study in reflexive ‘values in design’ (Knobel and Bowker 2011, Fish and Stark 2021) oriented to the public good, we apply the commonality metric to enhancing the diversity of recommendation by boosting a user population’s familiarity with underrepresented categories of music (or cultural content). To derive foundations for the values (diversity, universality, cultural citizenship) underlying this exercise, we turned to research that identifies the normative principles guiding public service media systems, and then attempted to translate them into recommender design.
Workshops have been held between recommendation engineers (Ferraro and Diaz) and SSH scholars (Ferreira and Born), generating mutual translation between CS and SSH spanning high-level methodological, epistemological and normative questions to concrete issues of design. Such sustained interdisciplinarity has enabled insights from SSH to be integrated with those from CS, and vice versa, with the goal of developing new models, methods and metrics for conceptualising, evaluating and designing novel recommendation paradigms. The team is committed to innovating in socially and culturally responsible design in recommender systems, and seeks to contribute to a growing body of scholarship developing public good rationales for digital media and machine learning systems.