An Approach to de-homogenizing recommender systems

Abstract

Recommender systems are all-pervasive on every online medium we interact with today. Yet, they impose a substantial problem of homogenization of users over time, leading to lack of visibility of the discriminated. This problem stems from the fundamental literature and traditional approaches around recommender systems and the primary focus on user-centric accuracy. In practice, this results in popularity bias, filter bubbles and "down the rabbit holes," where popular content in "your circle" continues to be recommended, and less popular content becomes discriminated against. And how does one measure popularity? Typically by some data metric of ratings: likes on a post, views on a video, the number of 5 star ratings. This ultimately results in locally optimal recommendations, where the users are content and comfortable with taking these results, but globally sub-optimal diversity, as the ecology of users grows more and more homogeneous over time. In this paper, I propose an approach to de-homogenizing recommender systems via an ecological values-based recommender system, as well as different metrics to evaluate on. This system encompasses both personalization, as well as diversification at both a user, and ecosystem of users, level. With this research, I demonstrate that this approach can introduce diversity in applicable ways, and reveals the weaknesses in current traditional models in tackling the problem of homogeneity. These insights can be used to guide future recommender systems and continue the conversation of developing more diverse and impactful recommendations.