Speaker
Christoph Best
(Google)
Description
To understand how well online advertising works (or if it works at all), we design geographically-based randomized experiments in which ads are shown to people in some regions, but not in other regions. The design of these regions, which form the randomizable units of the experiments, can be seen as a process of unsupervised learning about the underlying geographical structure of a country, based on how people move about. We introduce a suitable target function to evaluate the usefulness of a geographic decomposition and use spectral clustering to solve the learning problem.
Primary author
Christoph Best
(Google)
