Filtering the digital content reaching the user is a real added value for online platforms. A recommendation system, based on the user’s preferences, aims to maximize their satisfaction. By activating the brain’s reward system, which triggers dopamine release, it boosts engagement, making the user more likely to continue using the service. This prediction of attraction results from algorithmic calculations.
- Data Collection as a prerequisite
First of all, it is essential to highlight that recommendation systems rely on extensive data collection to discern individual preferences. Beyond the information users voluntarily provide when filling out their profile, a significant proportion of the knowledge comes from their digital behaviors. Every interaction – searches, connections, time spent on content, linguistic habits, etc. – contributes to enhancing the understanding of an individual. Michał Kosiński, computational psychologist, demonstrates this capability: “artificial intelligence can deduce someone’s personality and behavior from their digital traces, better than any human will.” “it is possible, from a minimum of 68 likes of an Internet user, to predict his/her skin color (at 95%), sexual orientation (88%), political beliefs (85%) and even determine whether or not the user’s parents are divorced”. Ultimately, this continuous flow of data feeds algorithmic models designed to predict and shape user choices.
- Focus on collaborative filtering
Among the various recommendation techniques, one of the most widely used is collaborative filtering. Collaborative filtering groups users with similar behaviors and preferences to predict their future interests. The basis of this filtering lies in a similarity of actions, the interest they have in a given content. This method is defined as neighborhood-based, and is described as including two possible approaches under the same logic:
- User-based collaborative filtering: based on the assumption that what has already appealed to one user will also appeal to his peer. The algorithm starts with user A and distinguishes that his search behavior is similar to that of B and C, neighbor. The system then highlights content liked by B and C to user A if he hasn’t yet seen it.
- Item-based collaborative filtering: the starting point is the proposed content. If it turns out that content 1 and content 2 have been liked by the same users, the algorithm deduces that 1 and 2 are neighboring, similar content; thus, if a user enjoyed one of them, the second will be offered to him.
Different ways of measuring similarity can be used for recommendation systems, notably cosine distance or Pearson’s correlation coefficient. Compared to content-based filtering, collaborative filtering allows for a greater diversity of content. However, it requires a very rich training corpus and cannot function with a small amount of data. That’s why new users face the problem of starting from scratch, so a latency time.
This approach, which pools the behaviors of a group to personalize the individual experience, finds a particular echo in the interpretation of theorist Bernard Harcourt. For Harcourt, this type of recommendation system is emblematic of what he calls the “logic of the digital Doppelgänger”.
- The Digital Doppelgänger: Bernard E. Harcourt’s Analysis
Bernard E. Harcourt is a multi-faceted jurist, professor, critical theorist and writer, specializing in issues of surveillance, power, punishment and political economy from a contemporary perspective. In his book Exposed. Desire and Disobedience in the Digital Age (2015), he analyzes the transformations of society and power brought about by technological advances, focusing on the omnipresence of surveillance and the erosion of privacy.
Chapter 5 is devoted to the concept of the “digital Doppelgänger”. The chapter begins by presenting the striking contrast between the difficulties of identification in the past and the disconcerting ease of identification in the digital age (through data). However, what characterized this new era is that the objective is no longer merely to identify an individual but to match them with their digital counterpart, to find for each of us our double, based on our consumption habits. The aim is to model the behavior of this digital double to anticipate our own behavior, and vice versa, in a process where the desires of each are reciprocally shaped.
He notes that while the old models focused on classifying individuals into groups to predict their behavior (actuarial methods), the new approach aims to establish perfect individual correspondences.
The word “Doppelgänger” means double in German, with an evil connotation. Bernard E. Harcourt observes that this digital double, both a tool of recommendation and control, challenges the liberal ideal of individual autonomy and dissolves the boundaries between privacy, commerce and surveillance. It has the potential to manipulate and influence individual subjectivity. Users become both subjects and objects, consumers of data and products of the digital system. In the end, this method is driven by several desires: to know, by controlling group behavior in order to ‘tame chance’; to classify, by arranging individuals in categories; and to insure, by reducing risk, mitigating responsibility and eliminating fault.
Laeticia Eschlimann
M2 Cyberjustice – Promotion 2024/2025
Sources :
- Bernard E. Harcourt, La société d’exposition : désir et désobéissance à l’ère numérique, La Couleur des idées, Le Seuil, 2020, 336p.
- La Grande table Idées d’Olivia Gesbert, France Culture, émission du 9 janvier 2020 : « Bernard Harcourt : La société d’exposition »
- Les nouvelles servitudes, La vie des idées, À propos de : Bernard E. Harcourt, La Société d’exposition, Désir et désobéissance à l’ère numérique
- J. Duportail, L’amour sous algorithme, éditions Goutte d’Or, 2019, p. 123, sur les recherches de Michal Kosinski, Computer-based personality judgments are more accurate than those made by humans, PNAS, 2015.
- What is collaborative filtering, IBM.
- Image by freepik
