(Talk 2 of 2)
Peter Marx, Vice President, Advanced Projects, GE Digital, Adjunct Professor, USC
Recent Advances in Post-Selection Statistical Inference
Robert Tibshirani, Professor of Biomedical Data Science, and Statistics Stanford University
Colloquium Series: Dr. Robert Schapire
Principal Researcher, Microsoft Research (NYC)
We study the general problem of how to learn through experience to make intelligent decisions. In this setting, called the contextual bandits problem, the learner must repeatedly decide which action to take in response to an observed context, and is then permitted to observe the received reward, but only for the chosen action. The goal is to learn through experience to behave nearly as well as the best policy (or decision rule) in some possibly very large and rich space of candidate policies. Previous approaches to this problem were all highly inefficient and often extremely complicated. In this work, we present a fast and simple algorithm that learns to behave as well as the best policy at a rate that is (almost) statistically optimal. Our approach assumes access to a kind of “oracle” (or subroutine) for classification learning problems which can be used to select policies; in practice, most off-the-shelf classification algorithms could be used for this purpose. Our algorithm makes very modest use of the oracle, which it calls far less than once per round, on average, a huge improvement over previous methods. These properties suggest this may be the most practical contextual bandits algorithm among all existing approaches that are provably effective for general policy classes.
This is joint work with Alekh Agarwal, Daniel Hsu, Satyen Kale, John Langford and Lihong Li.
The Data Science Institute's Centers for Data, Media & Society (formerly New Media) + Health Analytics will be holding a poster session, Wednesday, May 3rd from 2PM to 4PM in 407 Mudd featuring poster presentations and demonstrations related t
A panel of 4 hedge funds that are using Big Data and Machine Learning to invest, moderated by BAML PB Consulting.
Dr. Michael Winter
Senior Fellow, Advanced Technology
Pratt & Whitney, United Technologies Corporation
Systems Engineering: Imperatives, Definitions, Technology & Talent
The lecture will present the motivation, mechanics, and methodologies of model-based systems engineering as applied to product platforms and infrastructures that are often safety or operationally critical. Cyber-physical system-of-systems that combine both physics and controls form the basis of modern society. Application of systems engineering principles in an analytic context with focus on requirements, architecture, model-based development, and design flows will be presented as applied in an industrial context.
SVP, Data Science
Title: Estimating Causal Effect of Ads in a Real-Time Bidding Platform
Abstract: A real-time bidding platform responds to incoming ad-opportunities ("bid requests") by deciding whether or not to submit a bid and how much to bid. If the submitted bid wins, the user is shown an ad. Advertisers hope that ad-exposure leads to an increased likelihood of a desired action, such as a click or conversion (purchase, etc). So an important quantity that advertisers want to measure is the causal effect of advertising, namely, what is the response probability of an exposed user, compared with the counterfactual (un-observable) response-rate of the user if they were not exposed to the ad. In an ideal randomized test, the user is randomly assigned to test or control AFTER the submitted bid is won, and test users are served the ad in the normal way, while control users are not. While this is ideal from a statistical perspective, in practice this approach has the drawback that money spent by advertisers is wasted when a user is assigned to control. At Media Math we have developed a methodology for causal effect measurement where users are assigned to test or control BEFORE bid submission. One challenge here is that not all test-group users are exposed to an ad; only a winning bid results in ad exposure, and the winning population can have a significant bias. This talk will describe our approach to handle this and other challenges to ad impact measurement in this setting, and how we use MCMC Gibbs sampling to arrive at confidence intervals for ad-impact.