[Les Entretiens 1/5] Aurélie Jean: “The more a book is recommended, the more it is recommended, it’s a circle”

How do book recommendation algorithms work?

Not having put my hands under the hood, I can only imagine how they work. On an online book sales platform, a first machine learning algorithm (or machine learning) known as categorization, classifies users according to their static profile and their dynamic profile. The static profile contains all the information explicitly provided by the user when registering, such as their age, gender and reading preferences (science fiction, essays, comics, etc.). The dynamic profile contains all the information related to their behavior on the platform, such as the types of books they buy (pocket or paperback format, fiction or essay, type of literature – romantic comedy, classic, theater, etc. – , their purchase frequency, or even the type of book they stop at on the platform From this information, the categorization algorithm will construct classes of users according to statistical similarity in their static and profiles. -dynamic.

In return, a so-called recommendation algorithm will therefore recommend to a new user, whose class we will know from the projection of his data in the class space, a book that another user of his class liked with the the idea, in theory, that he will love it too. To this must be added a so-called explicit algorithm component, which will interfere with the results of the categorization algorithm by filtering the results. For example, if the user does not want to read novels, the algorithm is unlikely to suggest them.

Is the role of algorithms a simple reflection of trends, or does it actively influence readers' choices?

Certainly both! In practice, and explicitly, the books that sell the most will possibly be more highlighted on the platform, regardless of the class to which the user belongs or the reading preferences that they specified when registering. . To this, it must be added that there is a sort of circle that is created here: the more a book is recommended, the more it is seen and purchased, and therefore the more it is subsequently recommended. In this respect, the way in which recommendation algorithms are designed must take this effect into account by introducing statistical weight on the results of the suggestion or by directly influencing the functioning of the algorithm.

What are the advantages and disadvantages of recommendation algorithms?

The advantage is to have personalized recommendations based on user behavioral data by letting the algorithm capture strong and weak signals of their reading preferences. That being said, these algorithms – when they are moderately well designed or when they support a business model of impulse buying without a strategy of deep satisfaction – can prevent discovery by surprise or lock users into reading bubbles. In practice, the reader will mostly be recommended books that readers who are similar to them read.

Does sales analysis risk undermining editorial diversity?

This effect has always existed and is not specific to algorithmic recommendations. What I find interesting in algorithms is the analysis of reader preferences and behaviors to understand certain successes or failures. But publishers and booksellers should not take fewer risks with regard to algorithmic machinery that would guide readers' preferences and therefore future sales.

Do algorithms tend to favor large sellers or help the emergence of small ones?

Hard to say. I couldn't comment on this point. In practice, if an algorithm designed more simply highlights big sales without going further in the personalization of the recommendation, it is possible.

“Recommendation algorithms must exist in ideal complementarity with our super flesh-and-blood booksellers! »

How to preserve the uniqueness of bookstores?

Have the best booksellers, like mine! (Laugh.) The recommendation algorithm will never be better than a bookseller with his emotions, his practical and creative intelligence, his instinct and his intuition. Also by mastering all the components of intelligence that the machine does not master, they regularly surprise me by making suggestions which seem counterintuitive to me, but which in reality please me greatly! My bookseller, Fred from Librairie Tome 7 in Paris, is always right in recommending novels to me, even though I read few of them, not being naturally attracted to them.

What are the risks of a market homogenized by AI?

Prevent the existence of all forms of cultural change through the rise of a new literary style or by an author unknown to the general public. That being said, recommendation algorithms can help, but let's remember that these algorithms exist to operate platforms often run by non-booksellers, who seek to make a profit without any particular commitment to reading. (which booksellers can have).

Don’t the “reading bubbles” that algorithms generate risk, in the long term, extinguishing our curiosity and everything that pushes us towards the unknown?

This is why recommendation algorithms must exist in ideal complementarity with our super flesh-and-blood booksellers! This is also why radio, TV, podcast and other shows also have a bright future.