mar30 - STA 414/2104 Mar 30, 2010 Notes no class Thursday,...

Info iconThis preview shows pages 1–7. Sign up to view the full content.

View Full Document Right Arrow Icon
STA 414/2104 Mar 30, 2010 Notes I no class Thursday, April 2 I project due Thursday, April 16 before 2 pm I see http://www.utstat.utoronto.ca/reid/ sta414/414S10-html.doc for outline (Jan 5,7) I Take-home MT graded by Tuesday, April 6; pickup in SS 6003 1 / 25
Background image of page 1

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full DocumentRight Arrow Icon
STA 414/2104 Mar 30, 2010 The Netflix Grand Prize I “The BellKor Solution to the NGP”, Koren (August 2009) I “The BigChaos Solution to the NGP”, T¨oscher, Jahrer, Bell (September 2009) I “The Pragmatic Theory solution to the Netflix Grand Prize”, M. Piotte, M. Chabbert, (August 2009). I “The BellKor 2008 Solution to the Netflix Prize”, Bell, Koren, Volinsky I “All Together Now: A Perspective on the Netflix Prize”, Bell, Koren, Volinsky (2010) in Chance 23 , 24 – 29. I first four papers available from http://www.netflixprize.com//index 2 / 25
Background image of page 2
STA 414/2104 Mar 30, 2010 The data I approx 18,000 movies; nearly 500,000 users I total approx 100 million ratings, collected over seven years I ratings are 1 to 5 stars I collaborative filtering models : use models built on the training data to make individualized predictions I hold-out set of approx 4.2 million ratings I split into three subsets: probe set , quiz set , test set I prizemaster reports value of root-mean-squared-error (RMSE) for the quiz set on the leaderboard I winner determined by the first to improve RMSE on the test set by 10% beyond Netflix’s Cinematch system 3 / 25
Background image of page 3

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full DocumentRight Arrow Icon
STA 414/2104 Mar 30, 2010 The contest I Cinematch RMSE at the beginning of the contest: 0.9525 I benchmark to win: 90% of this: 0.8572 I winning entries: 0.8567 http://www.netflixprize.com//leaderboard 4 / 25
Background image of page 4
STA 414/2104 Mar 30, 2010 Some wrinkles I hold-out set (probe, quiz, test) are last nine movies rated by each user, (or fewer) I contains many more ratings by users that do not rate much I harder to predict I some users rated more than 10,000 movies I average number of ratings per user is 208 I 25% of users rated fewer than 50 movies I if we view ratings/users as a huge matrix r ui : 99% of the entries are missing I training set is T = { ( u , i ) | r ui is known } I R ( u ) : all the items rated by user u I R ( i ) : all the users who rated movie i I N ( u ) : all the items for which u provided a rating, even if the rating is unknown (i.e. qualifying set) 5 / 25
Background image of page 5

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full DocumentRight Arrow Icon
STA 414/2104 Mar 30, 2010 General strategy I avoid overfitting by regularizing parameter estimates I i.e. penalizing || θ || 2 (“ L 2 regularization”) I when a new tuning constant introduced, choose best RMSE on probe set over several runs I and keep that tuning constant fixed going forward I BigChaos introduced a second training run with the chosen tuning constant, using probe and training data I since various predictors will be aggregated, tuning could be chosen based on aggregation 6 / 25
Background image of page 6
Image of page 7
This is the end of the preview. Sign up to access the rest of the document.

Page1 / 25

mar30 - STA 414/2104 Mar 30, 2010 Notes no class Thursday,...

This preview shows document pages 1 - 7. Sign up to view the full document.

View Full Document Right Arrow Icon
Ask a homework question - tutors are online