eduzhai > Applied Sciences > Computer Science >

Parallel Experimentation in a Competitive Advertising Marketplace

  • KanKan
  • (0) Download
  • 20210425
  • Save

... pages left unread,continue reading

Document pages: 51 pages

Abstract: When multiple firms are simultaneously running experiments on a platform, thetreatment effects for one firm may depend on the experimentation policies ofothers. This paper presents a set of causal estimands that are relevant to suchan environment. We also present an experimental design that is suitable forfacilitating experimentation across multiple competitors in such anenvironment. Together, these can be used by a platform to run experiments "as aservice, " on behalf of its participating firms. We show that the causalestimands we develop are identified nonparametrically by the variation inducedby the design, and present two scalable estimators that help measure them intypical high-dimensional situations. We implement the design on the advertisingplatform of this http URL, an eCommerce company, which is also a publisher of digitalads in China. We discuss how the design is engineered within the platform sauction-driven ad-allocation system, which is typical of modern, digitaladvertising marketplaces. Finally, we present results from a parallelexperiment involving 16 advertisers and millions of this http URL users. These resultsshowcase the importance of accommodating a role for interactions acrossexperimenters and demonstrates the viability of the framework.

Please select stars to rate!


0 comments Sign in to leave a comment.

    Data loading, please wait...