1 / 28

Group Meeting – 3/16

Group Meeting – 3/16. Brian Ackerman. OrdRec. An Ordinal Model for Predicting Personalized Rating Distributions Yehuda Koren (Yahoo! Research) and Joe Sill Yehuda Koren won the Netflix Prize during his time at AT&T Voted best paper at RecSys’11. Road Map. Background Information

Download Presentation

Group Meeting – 3/16

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Group Meeting – 3/16 Brian Ackerman

  2. OrdRec • An Ordinal Model for Predicting Personalized Rating Distributions • Yehuda Koren (Yahoo! Research) and Joe Sill • Yehuda Koren won the Netflix Prize during his time at AT&T • Voted best paper at RecSys’11

  3. Road Map • Background Information • Recommender Systems • Learning to Rank • Stochastic Gradient Descent • OrdRec

  4. Recommender Systems – Setting • Input • Set of items (I) and a set of users (U) • Set of ratings (R) which take the form rui in R where rui is user u’s rating on item i • Output • Predictions for new user/item pairs

  5. Recommender Systems – SVD • Each user and item are represented by a feature vector of length k • E.g. Item A may be vector φA = [a1 a2 a3 … ak] • Imagine the features for items were fixed • E.g. items are movies and each feature is a genre such as comedy, drama, etc… • Features of the user vector are how well a user likes that feature

  6. Recommender Systems – SVD • Consider a movie (e.g. Die Hard) • Its feature vector may be φDH = [1 0 0] if the features are action, comedy, and drama • Maybe the user has the feature vector φU = [3.87 2.64 1.32] • We can try to predict a user’s rating using the dot product of these two vectors • r’U,DH(φU,φDH) = φU∙φDH = [1 0 0] ∙ [3.87 2.64 1.32] = 3.87

  7. Recommender Systems – SVD • The optimization goal for this approach is

  8. Road Map • Background Information • Recommender Systems • Learning to Rank • Stochastic Gradient Descent • OrdRec

  9. Learning to Rank – Architecture Training Data Learning System Model Test Data Predictions Ranking System

  10. Learning to Rank – Methods • Pointwise (OrdRec) • Each element of training data has a numerical or ordinal score (e.g. 1, 2, 3 or A, B, C) • Always a fixed scale (e.g. [1, 3] or A > B > C) • Pairwise • Pairs of elements have pairwise relationships (e.g. i > j, j > i, or i ~ j)

  11. Road Map • Background Information • Recommender Systems • Learning to Rank • Stochastic Gradient Descent • OrdRec

  12. Stochastic Gradient Descent • Learning method used for big data • Estimates the parameters of a model incrementally over a number of passes • Model is f(Θ) where Θ are the parameters • The SVD model only has two parameters and is r’u,i(Θ) = r’u,i(φu,φi)

  13. Stochastic Gradient Descent • Uses training data in the form of (x, r) pairs were x is an instance and r is the desired output • e.g. for instance x = (u, i), r = 3 • Goal is to learn Θthat best fit the model to minimize the error between the r’ values and r values

  14. Stochastic Gradient Descent • Update each parameter in Θ (θ in Θ) based on error (ε) with respect to model • E.g. updating all values in φuand φi • Use a small learning rate (λ) for gradual change • E.g. λ = .001

  15. Stochastic Gradient Descent • Update each parameter (θ) on each training pass over all instances • Δθ = λ * ε * (∂f(Θ) / ∂θ) • E.g. Δφu = λ * ε * (∂r’u,i(φu,φi) / ∂φu ) = λ * ε * φi • Example • ru,i = 3, r’u,i = 2.1, φu= .3, φi = .7 • ε = 3 – 2.1 = .9, λ = .001 • Δφu = .001 * .9 * .7 = .00147

  16. Road Map • Background Information • Recommender Systems • Learning to Rank • Stochastic Gradient Descent • OrdRec

  17. Motivation 1 Star = 1.0 2 Stars = 2.0 …

  18. Motivation A+ = ? B- = ? … A+ > B-

  19. Problem Setting • Input • Set of items (I) and a set of users (U) • Set of ratings (R) which take the form rui in R where rui is user u’s rating on item I, but all of these values fall into a relevance class (e.g. 1, 2, 3, …, S where S is the highest class)

  20. Problem Setting • Output • Rating predictions • Predicted distribution of ratings • Leads to recommendation confidence scores

  21. OrdRec Framework • Works with most existing recommender system methods • Framework assumes the internal scoring mechanism of an existing system is denoted as yui for user u and item i

  22. OrdRec Model • Likelihood function • Update based on …

  23. Thresholds • There are S – 1 ordered thresholds • t1 ≤ t2 ≤ … ≤ tS-1 • t1 is the only parameter • There is also a set of parameters β1, β2, …, βS-2 • tr+1 = tr + exp(βr)

  24. Thresholds • Assume we have ratings A, B, C • There are 2 thresholds (t1 and t2) as S = 3 • There is 1 threshold parameter (β1) exp(β1 ) t2 t1 C B A

  25. Methods • Given the parameters, the model generates an observed rating using a random process • zui ~ N(yui, 1) • The observed rating corresponds to the threshold where it falls • P(rui = r|Θ) = P(tr-1 < zui ≤ tr)

  26. Methods • Assuming t0 = -∞ and tS = ∞ • P(rui ≤ r|Θ) = P(zui ≤ tr) = Φ(tr– yui) • Φ is the CDF (cumulative distribution function) • New probability definition • P(rui = r|Θ) = P(rui ≤ r|Θ) – P(rui ≤ r – 1|Θ)

  27. OrdRec Model • Update Rule

  28. Predicting Ranking • Each user (u), and item pair (i > j) combination gets a vector (ΔPu(i,j)) • Each index represents and ordinal score (1, 2, …, S) • ΔPu(i,j)[r] = P(rui = r|Θ) - P(ruj = r|Θ) • Global weight vector (w) • ΔPu(i,j) and w yields a score

More Related