Content Creator Incentives in Recommender Systems

Title: Content Creator Incentives in Recommender Systems (via Zoom)

Abstract: When designing and evaluating content recommender systems such as Youtube or Netflix, it is common to treat the landscape of content available on the platform as static. However, in reality, the recommendation algorithm shapes the incentives of content creators and thus the content landscape, which can lead to unintended societal impacts. In this talk, we develop a high-dimensional, game-theoretic framework to analyze creator incentives, and we theoretically and empirically characterize downstream impacts on content diversity and clickbait prevalence. 
In our model, content creators compete for user consumption and strategically choose what D-dimensional content to create, while the platform optimizes for engagement. First, we characterize how content diversity depends on user preferences, production costs, and recommendation algorithm hyperparameters, under the simplifying assumption that engagement aligns with user value. Then, we turn to misaligned engagement metrics which reward both clickbait and quality: we characterize clickbait prevalence and downstream performance impacts. More broadly, our results highlight the need to carefully account for the endogeneity of the content landscape when evaluating recommendation algorithms. 
Based on joint work with Nikhil Garg, Nicole Immorlica, Brendan Lucier, and Jacob Steinhardt. 

Bio: Meena Jagadeesan is a 4th year PhD student in Computer Science at UC Berkeley, advised by Michael I. Jordan and Jacob Steinhardt. Her research investigates machine learning in digital marketplaces, with a focus on the interactions between ML algorithms and market competition.