Skip to content

Latest commit

 

History

History
40 lines (28 loc) · 1.79 KB

FAQs.md

File metadata and controls

40 lines (28 loc) · 1.79 KB

How did you obtain data from The New Yorker?

We currently run The New Yorker's Caption Contest. Their caption contest page says

Help us pick three finalists from last week’s contest by rating submissions. Vote now »

where "Vote now »" points to http://nextml.org/captioncontest

What are the algorithm differences?

The different algorithms choose which caption to display. The biggest difference between algorithms is "adaptive" vs "passive":

  • passive: these algorithms don't use any history or prior responses to choose which caption to display.
  • adaptive: these algorirthm choose which caption to display based on historyr/prior responses. They are designed with the goal of finding the funniest caption (meaning they have accurate estimates of the scores for funnier captions).

Adaptive algorithms include "LilUCB" and "KL-UCB". Passive algorithms include "RandomSampling" and "RoundRobin".

What questions do you ask?

  • How funny is this caption? (all contest)
  • Which of these two captions is funnier? (contest 508 and 509)
  • How original is this caption? (contest 560)

Columns in summary CSVs?

  • caption: The text that was displayed to the participant
  • funny, somewhat_funny and unfunny: The number of users that clicked that button
  • score: That caption's score. Calculated with 1*unfunny + 2*somewhat_funny + 3*funny
  • rank: That caption's rank. Captions with the same score have the same rank
  • count: The total number of times that caption has been rated by participants
  • contest: What's the contest where this caption has been displayed? The contest number increases by one every week
  • precision: The 95% confidence interval that the "true" score of that caption falls in score - precision and score + precision