Algodiv: Algorithmic recommandation and diversity of information on the web – ALGODIV

How to measure, assess and check the quality and adequacy of the choices made by web algorithms as they put forward such or such information? What are the options for academics in a context where more and more questions arise regarding the role of web computers in the organization of digital information and, more broadly, the life of Internet users? Could algorithms be "neutral"? What are the aims of those who introduce bias in these algorithms? What are the consequences of algorithmic governance on the autonomy and freedom of choice of Internet users?

Questions related to verifying, controling and governing the algorithms of the most important web platforms are pressing. This interdisciplinary project would bring a scientific contribution to this political, economic and cultural debate, by (i) empirically exploring the effects of algorithmic governance on informational diversity, (ii) building tools aiming at describing and measuring this impact and (iii) proposing alternatives to users. Starting from an endeavor of clarification of the various principles underlying the notion of "diversity", the project aims at conceiving metrics to appraise the effects of algorithms on Internet user behavior. By gathering computer and social scientists, Algodiv will aim at examining the challenges of the question of informational diversity in three distinct and key contexts of use of algorithms: structural authority measures from hypertext links (blogs, online media and Wikipedia), affinity measures from social web platforms (using Facebook [Algopol], Twitter and Instagram datasets, which are already available in the consortium), and personalized browsing traces from a unique dataset stemming from a partnership with the Groupe Le Monde (lemonde, telerama, rue89, etc.) providing exhaustive data on published, browsed and recommended content (browsing histories, audience measures).

These three main fields will be analyzed both from a computer science standpoint and from a sociological stance, in order to compare the rankings produced by the platforms and their algorithms, and the behavior of Interent users -- in some cases, ad hoc experimental settings will be developed to measure more finely the potential discrepancy between potential and actual informational and consumption diversity.