San Francisco Chapter of the American Statistical Association Seminar



William D Heavlin (, Google, Inc.


February 3rd, 2015  

4:30-6:00pm (4:30-5:00 networking/refreshment; 5:00-6:00 seminar)


Stanford University, Room 4205 of Scully Ctr. For Clinical Sciences Research building (CCSR)





 As American Idol attests, where one performance entertains, two performances compete. In 2011-2012 YouTube presented a few series of pairs of video performances (“slams”), and invited its viewers to vote for their favorite. The first-order goal is to determine better videos: the resulting rankings guided YouTube viewing suggestions. A secondary goal is to entertain the viewers voting: greater participation both increases data volume and enhances the YouTube brand.


Here we consider three statistical issues: (1) how to pair and play the videos, the tournament design problem; (2) how to rank the videos, the (primal) estimation problem; and (3) how to rank the acuity of the viewers, a (dual) estimation problem. Our approach likewise has three components: (a) a linearized objective function to schedule performance pairs, (b) a Bradley-Terry model to estimate performance quality; and (c) Fisher scores to assess voter acuity.



Direction for the location:


Directions to Stanford University can be found at


The seminar will take place in room 4205 (CCSR4205) of Scully Ctr. For Clinical Sciences Research building (CCSR).  Please refer to the campus map attached below for building location and parking options.  Interactive map is available at  Parking is free after 4:00pm at Stanford parking lots.


Meeting location direction