Menu

Eric Danan, Thibault Gajdos et Jean-Marc Tallon*

This article was originally published in the January 2021 edition of the 5 papers…in 5 minutes

PNG - 994.7 ko

Most Internet platforms that we use daily –– Facebook, Amazon, Spotify, LinkedIn, Trip Advisor and others –– offer us recommendations based on the massive volume of data that they collect. The exact way in which the recommendations are established varies from one platform to another, but they all work on a simple principle called collaborative filtering: the content that they recommend to you are those that other clients with similar tastes have enjoyed. The way in which this similarity is calculated is crucial to understanding the properties of the recommendation system algorithms.

In this article, Eric Danan, Thibault Gajdos and Jean-Marc Tallon take an axiomatic approach to the question. They suggest that the recommendation engines work on two basic, simple principles. The first is a principle of coherence: if the system has already registered that you have expressed a preference for article A over article B, then it recommends A rather than B, regardless of the views of other clients. The second principle, from the literature on social choice, is known as the Pareto principle: if all customers have expressed a preference for A over B then the system will recommend A to you rather than B (or will rank A higher than B). Danan, Gajdos and Tallon show first that the usual collaborative filter systems, based on the correlation of user profiles, satisfy neither of these two principles. Article B can be recommended rather than A even if all users with profiles similar to yours have preferred A over B. They then characterise all the rules that do satisfy the two principles mentioned. These are based not on correlation calculations among user profiles, but seek, rather, a combination of customer profiles that are close to yours. These profiles might be quite unlike yours (i.e. poorly correlated to yours), which is a marked difference from the existing systems. The abstract characterisation of these rules is a first step towards implementing them.

………………..

References

Original title of the article: Tailored Recommendations

Published in: 2020 Social Choice and Welfare, Springer Verlag, In press

Available at: https://hal.archives-ouvertes.fr/halshs-02973924/


* PSE Member

Credits : Shutterstock – metamorworks