Tailored recommendations
- PDF / 1,327,163 Bytes
- 20 Pages / 439.37 x 666.142 pts Page_size
- 8 Downloads / 224 Views
Tailored recommendations Eric Danan1 · Thibault Gajdos2 · Jean‑Marc Tallon3 Received: 9 June 2020 / Accepted: 5 October 2020 © Springer-Verlag GmbH Germany, part of Springer Nature 2020
Abstract Many popular internet platforms use so-called collaborative filtering systems to give personalized recommendations to their users, based on other users who provided similar ratings for some items. We propose a novel approach to such recommendation systems by viewing a recommendation as a way to extend an agent’s expressed preferences, which are typically incomplete, through some aggregate of other agents’ expressed preferences. These extension and aggregation requirements are expressed by an Acceptance and a Pareto principle, respectively. We characterize the recommendation systems satisfying these two principles and contrast them with collaborative filtering systems, which typically violate the Pareto principle.
1 Introduction The digitalization of our societies has enabled the personalization of advice to an extent never seen before. We now routinely receive recommendations or advertisements not for the “best” possible product, but, rather, for the one that is “best fit” for us. Netflix, for instance, adjusts the movies and shows it suggests to each user based on her profile, what she has seen and how much she appreciated it. Different users receive different recommendations. As for many other internet platforms using socalled “collaborative filtering” recommendation systems (Facebook, Twitter, Amazon, Spotify, Last.fm, LinkedIn, ...), the recommendation for a given user will hinge upon other users that are “similar” to her in one way or another. How this notion of
* Jean‑Marc Tallon jean‑[email protected] Eric Danan [email protected] Thibault Gajdos thibault.gajdos@univ‑amu.fr 1
CY Cergy Paris Université, CNRS, THEMA, 95000 Cergy, France
2
LPC, CNRS, Université d’Aix-Marseille, Marseille, France
3
Paris School of Economics, CNRS, Paris, France
13
Vol.:(0123456789)
E. Danan et al.
similarity between users is modeled is crucial for the properties of the recommendation system. The present paper approaches the issue of personalized, tailored recommendation from a normative perspective. Abstracting from the sophistication of actual recommendation systems, we explore the implications of imposing basic axiomatic principles in a simple formal framework. More specifically, we view a recommendation system as a way to “extend” an agent’s preferences to alternatives she has not yet rated, on the basis of some aggregate of other agents’ preferences. Preferences here encode the information available to the recommendation system—typically based on a user’s ratings or viewing behavior. Agents might actually be able to rank more alternatives than those they have rated, but such additional rankings are not modeled as the recommendation system cannot take them into account. We implicitly interpret the available information as being truthful, thereby leaving aside strategic considerations. We allow all agents’ prefer
Data Loading...
