Diversity in Network-Friendly Recommendations
Evangelia Tzimpimpaki, Thrasyvoulos Spyropoulos
TL;DR
The paper addresses the diversity loss that can arise when optimizing for network-friendly content recommendations. It introduces entropy-based diversity as a core objective and develops Diverse-NFR, a convex reformulation that is further tractable via a linear approximation to LP. The method yields high network gains with only modest reductions in diversity and remains compatible with existing fairness metrics, enabling joint consideration of reach, fairness, and network efficiency. The findings are validated conceptually and via simulation on real datasets, highlighting practical trade-offs and guiding deployment in content platforms that rely on caching and edge delivery.
Abstract
In recent years, the Internet has been dominated by content-rich platforms, employing recommendation systems to provide users with more appealing content (e.g., videos in YouTube, movies in Netflix). While traditional content recommendations are oblivious to network conditions, the paradigm of Network-Friendly Recommendations (NFR) has recently emerged, favoring content that improves network performance (e.g. cached near the user), while still being appealing to the user. However, NFR algorithms sometimes achieve their goal by shrinking the pool of content recommended to users. The undesirable side-effect is reduced content diversity, a phenomenon known as ``content/filter bubble''. This reduced diversity is problematic for both users, who are prevented from exploring a broader range of content, and content creators (e.g. YouTubers) whose content may be recommended less frequently, leading to perceived unfairness. In this paper, we first investigate - using real data and state-of-the-art NFR schemes - the extent of this phenomenon. We then formulate a ``Diverse-NFR'' optimization problem (i.e., network-friendly recommendations with - sufficient - content diversity), and through a series of transformation steps, we manage to reduce it to a linear program that can be solved fast and optimally. Our findings show that Diverse-NFR can achieve high network gains (comparable to non-diverse NFR) while maintaining diversity constraints. To our best knowledge, this is the first work that incorporates diversity issues into network-friendly recommendation algorithms.