User observations on usage of recommender systems to enable scaling of recommender systems experiments through automated and agentic means

Pursuant to article 27 DSA, providers of online platforms must clearly explain in their terms and conditions the main parameters used in their recommender systems, using plain and intelligible language. Moreover, they must explain options for users to modify or influence these main parameters. The main parameters that must be explained include (a) the most significant criteria used to determine suggested information, and (b) the reasons for the relative importance of these parameters. For platforms offering multiple options for recommender systems, they must provide a functionality allowing users to select and modify their preferred option at any time. Such a functionality should be directly and easily accessible from the section of the platform’s interface where the information is prioritized. In addition to the obligations established by article 27, providers of VLOPs and VLOSEs have an additional requirement: they must provide at least one recommender system option that is not based on profiling as defined in Regulation (EU) 2016/679 (“General Data Protection Regulation” or “GDPR”). These transparency requirements aim to give users more control and understanding of how information is presented to them on online platforms, particularly in relation to personalized recommendations. Against this background, this contract will explore how users engage with recommender systems across Very Large Online Platforms (VLOPs) and Very Large Online Search Engine (VLOSEs) through user surveys and user observations. DG CONNECT, together with the European Centre for Algorithmic Transparency at the Joint Research Centre (ECAT – JRC), is developing new methods to analyze recommender systems. These methods are developed across a wide range of representative recommendation surfaces across VLOPs and VLOSEs (both in terms of user experience – social media feeds, products listing, etc. - and types of contents – video, text, images, products, etc.). Therefore, the need arises to ground the understanding of such systems with real world data from observations of user interactions with recommender systems, confronting state of the art of the scientific literature with the reality of a representative panel of VLOPs and VLOSEs’ user base (in terms of relevant demographic variables such as age, gender, computer literacy, etc.).