Universität Potsdam Institut für Informatik Lehrstuhl Maschinelles Lernen. Recommendation. Tobias Scheffer

Transcription

1 Universität Potsdam Institut für Informatik Lehrstuhl Maschinelles Lernen Recommendation Tobias Scheffer

2 Recommendation Engines Recommendation of products, music, contacts,.. Based on user features, item features, and past transactions: sales, reviews, clicks, User-specific recommendations, no global ranking of items. Feedback loop: choice of recommendations influences available transaction and click data. 2

3 Netflix Prize Data analysis challenge, Netflix made rating data available: 500,000 users, 18,000 movies, 100 million ratings Challenge: predict ratings that were held back for evaluation; improve by 10% over Netflix s recommendation Award: $ 1 million. 3

4 Problem Setting Users U = {1,, m} Items X = {1,, m } Ratings Y = {(u 1, x 1, y 1 ), (u n, x n, y n )} Rating space y i Y E.g., Y = 1, +1, Y = {,., } Loss function l(y i, y j ) E.g., missing a good movie is bad but watching a terrible movie is worse. Find rating model: f θ : u, x y. 4

5 Problem Setting: Matrix Notation Users U = {1,, m} Items X = {1,, m } y 11 y 12 Ratings Y = y 21 y 23 y 33 Rating space y i Υ E.g.,Υ = item Loss function l(y i, y j ) user 1 user 2 user 3 1, +1,Υ = {,., } Incomplete matrix 5

6 Problem Setting Model f θ (u, x) Find model parameters that minimize risk θ = argmin θ l y, f θ u, x p u, x, y dxdudr As usual: p u, x, y is unknown minimize regularized empirical risk θ = argmin θ l y i, f θ u i, x i n i=1 + λω(θ) 6

7 Content-Based Recommendation Idea: User may like movies that are similar to other movies which they like. Requirement: attributes of items, e.g., Tags, Genre, Actors, Director, 7

8 Content-Based Recommendation Feature space for items E.g., Φ = comedy, action, year, dir tarantino, dir cameron T φ avatar = 0, 1, 2009, 0, 1 T 8

9 Content-Based Recommendation Users U = {1,, m} Items X = {1,, m } Ratings Y = {(u 1, x 1, y 1 ), (u n, x n, y n )} Rating space y i Y E.g.,Υ = 1, +1,Υ = {,., } Loss function l(y i, y j ) E.g., missing a good movie is bad but watching a terrible movie is worse. Feature function for items: φ: x R d Find rating model: f θ : u, x y. 9

10 Independent Learning Problems for Users Minimize regularized empirical risk θ = argmin θ l y i, f θ u i, x i n i=1 One model per user: f θu x Υ One learning problem per user: θ u = argmin θu l y i, f θu x i i:u i =u + λω(θ) + λω(θ u ) 10

11 Independent Learning Problems for Users One learning problem per user: u: θ u = argmin θu l y i, f θu x i i:u i =u + λω(θ u ) Use any model class and learning mechanism; e.g., f θu x i = φ x i T θ u Logistic loss + l 2 regularization: logistic regression Hinge loss + l 2 regularization: SVM Squared loss + l 2 regularization: ridge regression 11

12 Independent Learning Problems for Users Obvious disadvantages of independent problems: Commonalities of users are not exploited, User does not benefit from ratings given by other users, Poor recommendations for users who gave few ratings. Rather use joint prediction model: Recommendations for each user should benefit from other osers ratings. 12

13 Independent Learning Problems θ 1 Parameter vectors of independent prediction models for users Regularizer θ 2 13

14 Joint Learning Problem θ 1 Regularizer Parameter vectors of independent prediction models for users θ 2 14

15 Joint Learning Problem Standard l 2 regularization follows from the assumption that model parameters are governed by normal distribution with mean vector zero. Instead assume that there is a non-zero population mean vector. 15

16 Joint Learning Problem 0 θ Σ θ 1 θ 2 θ m Graphical model of hierarchical prior 16

17 Joint Learning Problem Population mean vector θ ~N 0, 1 I λ User-specific mean vector: θ u ~N θ, 1 λ I Substitution: θ u = θ + θ u ; now θ and θ u have mean vector zero. -Log-prior = regularizer Ω θ + θ u = λ θ 2 + λ θu 2 17

18 Joint Learning Problem Joint optimization problem: min θ 1,,θ m,θ u i:u i =u l y i, f θu +θ x i + λω θ u + λ Ω(θ ) Parameters θ u are independent, θ is shared. Hence, θ u are coupled. θ u = θ + θ u Coupling strength Global regularization 18

19 Discussion Each user benefits from other users ratings. Does not take into account that users have different tastes. Two sci-fi fans may have similar preferences, but a horror-movie fan and a romantic-comedy fan do not. Idea: look at ratings to determine how similar users are. 19

20 Collaborative Filtering Idea: People like items that are liked by people who have similar preferences. People who give similar ratings to items probably have similar preferences. This is independent of item features. 20

21 Collaborative Filtering Users U = {1,, m} Items X = {1,, m } Ratings Y = {(u 1, x 1, y 1 ), (u n, x n, y n )} Rating space y i Υ E.g.,Υ = 1, +1,Υ = {,., } Loss function l(y i, y j ) Find rating model: f θ : u, x y. 21

22 Collaborative Filtering by Nearest Neighbor Define distance function on users: d(u, u ) Predicted rating: f θ u, x = k nearest neighbors u i of u 1 k y u i,x Predicted rating is the average rating of the k nearest neighbors in terms of d u, u. No learning involved. Performance hinges on d(u, u ). 22

23 Collaborative Filtering by Nearest Neighbor Define distance function on users: d u, u = 1 m m x=1 y u,x, y u,x 2 Euclidean distance between ratings for all items. Skip items that have not been rated by both users. 23

24 Extensions Normalize ratings (subtract mean rating of user, divide by user s standard deviation) Weight influence of neighbors by inverse of distance. Weight influence of neighbors with number of jointly rated items. 1 d(u, u i ) y u i,x f θ u, x = k nearest neighbors u i of u k nearest neighbors u i of u 1 d(u, u i ) 24

25 Matrix Zombiland Titanic Death Proof Collaborative Filtering: Example Y = Alice Bob Carol How much would Alice enjoy Zombiland? 25

26 M Z T D Collaborative Filtering: Example Y = d u, u = 1 m m x=1 Alice Bob Carol y u,x, y u,x 2 d A, B = d A, C = d B, C = 26

27 M Z T D Collaborative Filtering: Example Y = d u, u = 1 m m x=1 Alice Bob Carol y u,x, y u,x 2 d A, B = 2.9 d A, C = 1 d B, C =

28 M Z T D Collaborative Filtering: Example Y = f θ A, Z = Alice Bob Carol 1 2 nearest d(a,u i ) y u i,z neighbors u i of A 1 2 nearest d(a,u i ) neighbors u i of A = 28

29 M Z T D Collaborative Filtering: Example Y = f θ A, Z = Alice Bob Carol 1 2 nearest d(a,u i ) y u i,z neighbors u i of A 1 2 nearest d(a,u i ) neighbors u i of A =

30 Collaborative Filtering: Discussion K nearest neigbor and similar methods are called memory-based approaches. There are no model parameters, no optimization criterion is being optimized. Each prediction reuqires an iteration over all training instances impractical! Better to train a model by minimizing an appropriate loss function over a space of model parameter, then use model to make predictions quickly. 30

31 Latent Features Idea: Instead of ad-hoc definition of distance between users, learn features that actually represent preferences. If, for every user u, we had a feature vector ψ u that describes their preferences, Then we could learn parameters θ x for item x such that θ x T ψ u quantifies how much u enjoys x. 31

32 Latent Features Or, turned around, If, for every item x we had a feature vector φ x that characterizes its properties, We could learn parameters θ u such that θ u T φ x quantifies how much u enjoys x. In practice some user attributes ψ u and item attributes φ x are usually available, but they are insufficient to understand u s preferences and x s relevant properties. 32

33 Latent Features Idea: construct user attributes ψ u and item attributes φ x such that ratings in training data can be predicted accurately. Decision function: f Ψ,Φ u, x = ψ u T φ x Prediction is product of user preferences and item properties. Model parameters: Matrix Ψ of user features ψ u for all users, Matrix Φ of item features φ x for all items. 33

34 Latent Features Optimization criterion: Ψ, Φ = argminψ,φ x,u + λ ψ u 2 u l(yu,x, f Ψ,Φ u, x ) + φ x 2 x Feature vectors of all users and all Items are regularized 34

35 Latent Features Both item and user features are the solution of an optimization problem. Number of features k has to be set. Meaning of the features is not pre-determined. Sometimes they turn out to be interpretable. 35

36 Matrix Factorization Decision function: f Ψ,Φ u, x In matrix notation: Matrix elements: y 11 y 1m y m1 y mm = ψ u T φ x Y Ψ,Φ = ΨΦ T = ψ 11 ψ 1k ψ m1 ψ mk φ 11 φ m 1 φ 1k φ m k 36

37 M Z T D Matrix Factorization Decision function in matrix notation: Predicted rating of Matrix for Alice y 11 y 1m y m1 y mm = ψ 11 ψ 1k ψ m1 Alice Bob Carol ψ mk φ 11 φ m 1 φ 1k φ m k Latent features of Alice Latent features of Matrix 37

38 M Z T D Matrix Factorization Decision function in matrix notation: Latent features of Alice y 11 y 1m y m1 y mm = ψ 11 ψ 1k ψ m1 Alice Bob Carol ψ mk φ 11 φ m 1 φ 1k φ m k Latent features of Carol Latent features of Matrix Latent features of Death Proof 38

39 Matrix Factorization Optimization criterion: Ψ, Φ = argminψ,φ x,u + λ Ψ 2 + Φ 2 Criterion is not convex: l(yu,x, f Ψ,Φ u, x ) For instance, multiplying all feature vectors with -1 gives an equally good solution: f Ψ,Φ u, x = ψ u T φ x = ( ψ u T )( φ x ) Limiting the number of latent features to k restricts the rank of matrix Y. 39

40 Matrix Factorization Optimization criterion: Ψ, Φ = argminψ,φ Optimization by x,u + λ Ψ 2 + Φ 2 l(yx,u, Stochastic gradient descent or Alternating least squares. f Ψ,Φ u, x ) 40

41 Matrix Factorization by Stochastic Gradient Descent Iterate through ratings y u,x in training sample Let ψ u ψ u α f Ψ,Φ(u,x) ψ u Let φ x φ x α f Ψ,Φ(u,x) φ x Until convergence. Requires differentiable loss function; e.g., squared loss, 41

42 Matrix Factorization by Alternating Least Squares For squared loss and parallel architectures. Alternate between 2 optimization processes: Keep Φ fixed, optimize ψ u in parallel for all u. Keep Ψ fixed, optimize φ x in parallel for all x. 42

43 Matrix Factorization by Alternating Least Squares For squared loss and parallel architectures. Alternate between 2 optimization processes: Keep Φ fixed, optimize ψ u in parallel for all u. Keep Ψ fixed, optimize φ x in parallel for all x. Optimization criterion for Ψ: ψ u = argmin ψu y u y u 2 λ ψ u 2 = argmin ψu y u ψ u T Φ T 2 λ ψ u 2 y u1 y um = ψ u1 ψ uk φ 11 φ m 1 φ 1k φ m k 43

44 Matrix Factorization by Alternating Least Squares For squared loss and parallel architectures. Alternate between 2 optimization processes: Keep Φ fixed, optimize ψ u in parallel for all u. Keep Ψ fixed, optimize φ x in parallel for all x. Optimization criterion for Ψ: ψ u = argmin ψu y u ψ u T Φ T 2 λ ψ u 2 ψ u = ΦΦ T + λi 1 Φy u 44

45 Matrix Factorization by Alternating Least Squares For squared loss and parallel architectures. Alternate between 2 optimization processes: Keep Φ fixed, optimize ψ u in parallel for all u. Keep Ψ fixed, optimize φ x in parallel for all x. Optimization criterion for Ψ: ψ u = argmin ψu y u ψ u T Φ T 2 λ ψ u 2 ψ u = ΦΦ T + λi 1 Φy u Optimization criterion for Φ: φ x = argmin φx y x φ x T Ψ T 2 λ φ x 2 φ x = ΨΨ T + λi 1 Ψy x 45

46 Matrix Factorization by Alternating Least Squares For squared loss and parallel architectures. Initialize Ψ, Φ randomly. Repeat until convergence: Keep Ψ fixed, for all u in parallel: ψ u = ΦΦ T + λi 1 Φy u Keep Φ fixed, for all x in parallel: φ u = ΨΨ T + λi 1 Ψy x 46

47 Extensions: Biases Some users just give optimistic or pessimistic ratings; some items are hyped. Decision function: f Ψ,Φ,Bu,B x u, x = b u + b x + ψ u T φ x Optimization criterion: Ψ, Φ, B u, B x = argmin Ψ,Φ l(y x,u, f Ψ,Φ,Bu,B x u, x ) x,u + λ Ψ 2 + Φ 2 + B u 2 + Bx 2 47

48 Extensions: Explicit Features Often, explicit user and item features are available. Concatenate vectors ψ u and φ x ; explicit features are fixed, latent features are free paremeters. 48

49 Extensions: Temporal Dynamics How much a user likes an item depends on the point in time when the rating takes place. f Ψ,Φ,Bu,B x,t u, x = b u t + b x (t) + ψ u t T φ x 49

50 Summary Purely content-based recommendation: users don t benefit from other users ratings. Collaborative filtering by nearest neighbors: fixed definition of similarity of users. No model parameters, no learning. Has to iterate over data to make recommendation. Latent factor models, matrix factorization: user preferences and item properties are free parameters, optimized to minimized discrepancy between inferred and actual ratings. 50