How Audit Methods Impact Our Understanding of YouTube’s Recommendation Systems

Abstract

Computational audits of social media websites have generated data that forms the basis of our understanding of the problematic behaviors of algorithmic recommendation systems. Focusing on YouTube, this paper demonstrates that conducting audits to make specific inferences about the underlying content recommendation system is more methodologically challenging than one might expect. Obtaining scientifically valid results requires considering many methodological decisions, and each of these decisions incurs costs. For example, should an auditor use logged-in YouTube accounts while gathering recommendations to ensure more accurate inferences from the collected data? We systematically explore the impact of this and many other decisions and make important discoveries about the methodological choices that impact YouTube’s recommendations. Assessed together, our research suggests auditing configurations that can be used by researchers and auditors to reduce economic and computing costs, without sacrificing inference quality and accuracy.