A Bayesian Model to Predict Content Creation with Two-Sided Peer Influence in Content Platforms

Abstract

While prior research has studied the motivations of individuals to consume content on social media platforms, limited work exists on how contributors are motivated to create content. We examine the role of peer influence in content production on YouTube, where content creators are competing for attention. Given that content creation efforts are driven not only by their personal preferences, but also by the content creation decisions of others in the network neighbors, we develop a new method to analyze discrete choice decisions (such as creating content or not) in a networked environment with panel data. We face a novel set of big data challenges, i.e., both statistical and quantitative, in estimating peer influence. We face computational challenges in that we cannot reasonably estimate peer influence over the entire YouTube network, which has billions of nodes. We employ graph sampling methods to address this issue. Identification of social influence in large-scale social networks such as YouTube is difficult due to the interdependence in decisions of users, correlations between the video's observable and unobservable characteristics and attributes over time. These patterns cannot be modeled with existing autocorrelation models. We design a new method, the Network Auto-Probit Model with Fixed Effects (NAFE), to identify peer influence among content creators on YouTube. Implications for research and practice are also discussed.

Keywords: social media, peer influence, auto-probit, Bayesian methods, big data analytics

JEL Classification: C11, C15, C33, C35, C53, L86, M20, M31