Prediction Accuracy And Autonomy : Assessing how recommender systems objectives can align with user autonomy

Abstract

Entertainment recommender systems have been criticised by journalists and tech-industry insiders for undermining individuals’ autonomy. These systems might exercise unwanted control over peoples’ lives, not through coercion but rather through distraction. In this thesis we adopt an interdisciplinary framework to explore how the design of recommendation systems for entertainment services can align with the individual right to autonomy. First, we assess design objectives by doing a corpus analysis on 1,883 scientific articles on entertainment recommender systems. We then carry out a qualitative survey of psychological literature and connect findings on self-regulation, sense of agency and habits to the autonomy of users. We also survey relevant literature on user-centred interaction design to relate the notion of user autonomy with user value. Finally, we focus on the specific use-case of YouTube’s recommender system and propose design changes aimed at better aligning service provider objectives with users’ objectives. We conclude that because of an intention-behaviour gap, users’ behaviour is an inaccurate reflection of users’ intentions. Because of this, only analysing behavioural data undermines users’ autonomy. Many current recommender systems, including YouTube’s, use behavioural data since the data is easy to collect and often maximise service providers’ goals. We propose both corrective and preventive solutions to this problem. The corrective solutions focus on offering users more customisability. The preventive solutions focus on ways to gather more data that better correspond to users’ intentions. Higher user customisability can provide user data that can be expected to correspond relatively well to users’ intention.