A0633
Title: Network experimentation at Meta
Authors: Liang Shi - Meta (United States) [presenting]
Abstract: Randomized experiments, or A/B tests, are nowadays common practices in tech companies such as Meta and Netflix to measure the causal effects of product changes. One big challenge that experimenters often face is the presence of network effects, which violates one of the common assumptions employed in analyzing these experiments, the stable unit treatment value assumption (SUTVA), i.e., an experimental unit's response to an intervention only depends on its own treatment status. Some common frameworks are presented that are deployed at Meta to tackle this challenge from both experiment design and analysis perspectives, and share some practical advice when applying these frameworks in real-world use cases.
