20th AIAI 2024, 27 - 30 June 2024, Corfu, Greece

Controlling Popularity Bias in Sequential Recommendation Models

Brandon Weaver, Melody Moh, Teng-Sheng Moh

Abstract:

  Society today has witnessed a rapid increase in the prevalence of social media and popular content delivery websites such as Netflix and Spotify.As a result, through the way that people consume news and media, we are transitioning from a static media delivery model to a dynamic, personalized system which many are adapting and even enjoying the resulted changes. Personalized recommendations are mostly made with the help of the machine learning models embedded in recommendation systems (RS). It is important to examine these systems that are becoming very relevant in our daily lives, and to identify how we can either control or at least mitigate any harmful effects caused by these new systems. One of those prominent issues within RS is popularity bias, which stems from the nature of machine learning, which often prioritizes popularity over novelty. That is, machine learning methods tend to prioritizes being most correct rather than trying to find a truly fitting recommendation. Popularity bias is a main cause of echo chambers within the current media landscape, which unfortunately has led to less critical thinking and more divisiveness our communities. To counter this issue, we present a novel methodology that combines two existing methods proposed for sequential RS, aiming to reduce popularity bias without overly punish popular items.The result is a weighted RS that shows the promising ability to control the amount of popularity bias. The proposed system may be applied to multi-model hybrid RS, which would achieve personalized recommendations that support individual weighted preference between popularity and novelty.  

*** Title, author list and abstract as seen in the Camera-Ready version of the paper that was provided to Conference Committee. Small changes that may have occurred during processing by Springer may not appear in this window.