NFR-Induced Diversity Reduction Conjecture

Determine whether Network-Friendly Recommendations (NFR) algorithms that seek to reduce network cost do so by consistently reallocating recommendation probability mass from non-cached items to cached items whenever sufficiently relevant alternatives exist, thereby concentrating recommendation probabilities on a smaller subset of content and reducing overall diversity in the recommendation system.

Background

Network-Friendly Recommendations (NFR) aim to improve network efficiency by biasing recommendations toward content that is cheaper to deliver, such as items cached near users. While prior NFR work has addressed individual user relevance and global intrusiveness, the potential reduction in content diversity—leading to so-called content/filter bubbles—had not been explicitly measured or incorporated into NFR algorithms.

The conjecture posits a mechanism by which NFR may reduce diversity: by shifting recommendation probability mass from non-cached to cached items whenever they are sufficiently relevant, NFR could systematically concentrate recommendations on a smaller set of items, potentially disadvantaging broader content exploration and content creators outside the cache.

References

Our conjecture is that, to reduce network cost, an NFR algorithm will replace some recommendation probability mass from items outside the cache, with items inside the cache, at every opportunity where sufficiently relevant alternatives (e.g. satisfying (i)) exist. By doing this for most users and most items, the majority of the recommendation probability mass will become concentrated around a much smaller pool of content, signaling decreased diversity in the RS.

— Diversity in Network-Friendly Recommendations  (2411.00601 - Tzimpimpaki et al., 2024) in Introduction — The problem: Diversity in NFR (Section I)