Exposure diversity in music recommender systems

During this five month project with Fabien Tarissan, I studied the diversity of the recommendations made by music recommender systems. We built upon a diversity measure in heterogeneous networks and applied it to study the algorithm from Collaborative Filtering For Implicit Datasets.

UPDATE: The extended version of our work was accepted in Applied Network Science

• Paper : pdf

UPDATE: I presented our work at the CNA 2021 conference in Madrid. Here is the related material:

• Extended abstract: pdf
• Presentation slides: pdf

Are the music recommendations locking you in a particular genre ?

Usually, recommendations are known to expose users to a greater diversity of items than what they would have searched themselves. However, the studies which came to such results only considered one aspect of diversity, namely the variety of reached categories.

With our new approach, we were able to show that despite increasing exposure variety for all users, recommendations significanlty reduce user's exposure balance. In other words, the recommendations are often strongly biased towards one or two musical categories (the user's favorites) and sometimes recommend unrelated items almost "by mistake".

For more information, here is our submission to RecSys 2021, along with the code.