Large-scale Collaborative Prediction Using a Nonparametric Random Effects Model

Size: px

Start display at page:

Download "Large-scale Collaborative Prediction Using a Nonparametric Random Effects Model"

Shon Melton
5 years ago
Views:

1 Large-scale Collaborative Prediction Using a Nonparametric Random Effects Model Kai Yu Joint work with John Lafferty and Shenghuo Zhu NEC Laboratories America, Carnegie Mellon University First Prev Page 1 Next Last Go Back Full Screen Close Quit

2 Learning multiple tasks For input vector z j, and its outputs Y ij (tasks), the standard regression model is under various conditions Y ij = µ + m i (z j ) + ɛ ij, where ɛ ij iid N(0, σ 2 ), i = 1,..., M, and j = 1,..., N. First Prev Page 2 Next Last Go Back Full Screen Close Quit

3 A kernel approach to multi-task learning To model the dependency between tasks, a hierarchical Bayesian approach may assume where m i iid GP(0, Σ) Σ(z j, z j ) 0 is a shared covariance function among inputs; Many multi-task learning approaches are similar to this. a a ICML-09 tutorial, Tresp & Yu First Prev Page 3 Next Last Go Back Full Screen Close Quit

4 Using task-specific features Assuming task-specific features x are available, a more flexible approach is to model the data jointly, as Y ij = µ + m(x i, z j ) + ɛ ij, where ɛ ij iid N(0, σ 2 ), m ij = m(x i, z j ) is a relational function. First Prev Page 4 Next Last Go Back Full Screen Close Quit

5 A nonparametric kernel-based approach Assume the relational function follows a m GP(0, Ω Σ) where Ω(x i, x i ) is a covariance function on tasks; Σ(z j, z j ) is a covariance function on inputs; any sub matrix follows m N(0, Ω Σ) Cov(m ij, m i j ) = Ω ii Σ jj ; If Ω = δ, the prior reduces to m iid i GP(0, Σ). a Yu et al., 2007; Bonilla et al., 2008 First Prev Page 5 Next Last Go Back Full Screen Close Quit

6 The collaborative prediction problem This essentially a multi-task learning problem with task features; Matrix factorization using additional row/column attributes; The formulation applies to many relational prediction problems. First Prev Page 6 Next Last Go Back Full Screen Close Quit

7 Challenges to the kernel approach Computation: the cost O(M 3 N 3 ) is prohibitive. Netflix data: M = 480, 189 and N = 17, 770. Dependent noise : when Y ij cannot be fully explained by the predictors x i and z j, the conditional independence assumption is invalid, which means, p(y m, x, z) i,j p(y ij m, x i, z j ) User and movie features are weak predictors; The relational observations Y ij alone are informative to each other. First Prev Page 7 Next Last Go Back Full Screen Close Quit

8 This work Novel multi-task model using both input and task attributes; Nonparametric random effects to resolve dependent noises ; Efficient algorithm for large-scale collaborative prediction problems. First Prev Page 8 Next Last Go Back Full Screen Close Quit

9 Nonparametric random effects Y ij = µ + m ij + f ij + ɛ ij, m(x i, z j ): a function depending on known attributes; f ij : random effects for dependent noises ; modeling dependency in observations with repeated structures. Let f ij be nonparametric: dimensionality increases with data size; nonparametric matrix factorization First Prev Page 9 Next Last Go Back Full Screen Close Quit

10 Efficiency considerations in modeling To save computation, we absorb ɛ into f and obtain Y ij = µ + m ij + f ij, Introduce a special generative process for m and f... m, f, Ω 0 (x i, x i ), Σ 0 (z j, z j ), First Prev Page 10 Next Last Go Back Full Screen Close Quit

11 The row-wise generative model First Prev Page 11 Next Last Go Back Full Screen Close Quit

12 The row-wise generative model First Prev Page 12 Next Last Go Back Full Screen Close Quit

13 The row-wise generative model First Prev Page 13 Next Last Go Back Full Screen Close Quit

14 The row-wise generative model First Prev Page 14 Next Last Go Back Full Screen Close Quit

15 The row-wise generative model First Prev Page 15 Next Last Go Back Full Screen Close Quit

16 The row-wise generative model First Prev Page 16 Next Last Go Back Full Screen Close Quit

17 The row-wise generative model First Prev Page 17 Next Last Go Back Full Screen Close Quit

18 The row-wise generative model First Prev Page 18 Next Last Go Back Full Screen Close Quit

19 The column-wise generative model First Prev Page 19 Next Last Go Back Full Screen Close Quit

20 Two generative models Y ij = µ + m ij + f ij, column-wise model: Ω IWP(κ, Ω 0 + τδ), m GP(0, Ω Σ 0 ), f iid j GP(0, λω), row-wise model: Σ IWP(κ, Σ 0 + λδ), m GP(0, Ω 0 Σ), f iid i GP(0, τσ), First Prev Page 20 Next Last Go Back Full Screen Close Quit

21 Two generative models are equivalent Both models lead to the same matrix-variate Student-t process Y MTP ( κ, 0, (Ω 0 + τδ), (Σ 0 + λδ) ), The model learns both Ω and Σ simultaneously; Sometimes one model is computationally cheaper than the other. First Prev Page 21 Next Last Go Back Full Screen Close Quit

22 An idea of large-scale modeling First Prev Page 22 Next Last Go Back Full Screen Close Quit

23 Modeling large-scale data If Ω 0 (x i, x i ) = φ(x i ), φ(x i ), m GP(0, Ω 0 Σ) implies m ij = φ(x i ), β j Without loss of generality, let Ω 0 (x i, x i ) = p 1 2 xi, p 1 2 xi, x i R p. On a finite observational matrix Y R M N, M N, the row-wise model becomes Σ IW(κ, Σ 0 + λi N ), β N(0, I p Σ), where β is p N random matrix. Y i N(β x i, τσ), i = 1,..., M First Prev Page 23 Next Last Go Back Full Screen Close Quit

24 Approximate Inference - EM E-step: compute the sufficient statistics {υ i, C i } for the posterior of Y i given the current β and Σ: Q(Y) = M p(y i Y Oi, β, Σ) = i=1 M N(Y i υ i, C i ), i=1 M-step: optimize β and Σ: β, Σ = arg min β,σ and then let β β, Σ Σ. { } E Q(Y) [ log p(y, β, Σ θ)] First Prev Page 24 Next Last Go Back Full Screen Close Quit

25 Some notation Let J i {1,..., N} be the index set of the N i observed elements in the row Y i ; Σ [:,Ji ] R N N i is the matrix obtained by keeping the columns of Σ indexed by J i ; Σ [Ji,J i ] R N i N i is obtained from Σ [:,Ji ] by further keeping only the rows indexed by J i ; Similarly, we can define Σ [Ji,:], Y [i,ji ] and m [i,ji ]. First Prev Page 25 Next Last Go Back Full Screen Close Quit

26 The EM algorithm E-step: for i = 1,..., M m i = β x i, υ i = m i + Σ [:,Ji ]Σ 1 [J i,j i ](Y [i,ji ] m [i,ji ]), C i = τσ τσ [:,Ji ]Σ 1 [J i,j i ]Σ [Ji,:]. M-step: β = (x x + τi p ) 1 x υ, [ M ] τ 1 i=1 Σ (C i + υ i υ i ) υ x(x x + τi p ) 1 x υ + Σ 0 + λi N = M + 2N + p + κ On Netflix, each EM iteration takes several thousands of hours. First Prev Page 26 Next Last Go Back Full Screen Close Quit

27 Fast implementation Let U i R N N i be a column selection operator, such that Σ [:,Ji ] = ΣU i and Σ [Ji,J i ] = U i ΣU i. The M-step only needs C = M i=1 C i + υ υ and υ x from the previous E-step. To obtain them, it s unnecessary to compute υ i and C i. For example, M C i = i=1 = M i=1 M i=1 ( ) τσ τσ [:,Ji]Σ 1 [J Σ i,j i] [J i,:] ( ) τσ τσu i Σ 1 [J U i,j i] i Σ = τmσ τσ ( M i=1 U i Σ 1 [J i,j i] U i Similar tricks can be applied to υ υ and υ x. Time for each iteration is reduced from thousands of hours to 5 hours only. ) Σ First Prev Page 27 Next Last Go Back Full Screen Close Quit

28 EachMovie Data users, 1648 movies; 2, 811, 718 numeric ratings Y ij {1,..., 6}; 97.17% of the elements are missing; Use 80% ratings of each user for training and the rest for testing; This random selection is repeated 10 times independently. First Prev Page 28 Next Last Go Back Full Screen Close Quit

29 Compared methods User Mean & Movie Mean: prediction by the empirical mean; FMMMF: fast max-margin matrix factorization a ; PPCA: probabilistic principal component analysis b ; BSRM: Bayesian stochastic relational model c, BSRM-1 uses no additional user/movie attributes d ; NREM: Nonparametric random effects model, NREM-1 uses no additional attributes. a Rennie & Srebro (2005). b Tipping & Bishop (1999). c Zhu, Yu, & Gong (2009). d Top 20 eigenvectors from of binary matrix indicating if a rating is observed or not. First Prev Page 29 Next Last Go Back Full Screen Close Quit

30 Results on EachMovie TABLE: Prediction Error on EachMovie Data Method RMSE Standard Error Run Time (hours) User Mean Movie Mean FMMMF PPCA BSRM BSRM NREM NREM First Prev Page 30 Next Last Go Back Full Screen Close Quit

31 Netflix Data 100, 480, 507 ratings from 480, 189 users on 17, 770 movies; Y ij {1, 2, 3, 4, 5}; A set of validation data contain 1, 408, 395 ratings; Therefore there are 98.81% of elements missing in the rating matrix; The test set includes 2, 817, 131 ratings; First Prev Page 31 Next Last Go Back Full Screen Close Quit

32 Compared methods In addition to those compared in EachMovie experiment, there are several other methods: SVD: a method almost the same as FMMMF, using a gradient-based method for optimization a. RBM: Restricted Boltzmann Machine using contrast divergence b. PMF and BPMF: probabilistic matrix factorization c, and its Bayesian version d. PMF-VB: probabilistic matrix factorization using a variational Bayes method for inference e. a Kurucz, Benczur, & Csalogany, (2007). b Salakhutdinov, Mnih & Hinton (2007). c Salakhutdinov & Mnih (2008b). d Salakhutdinov & Mnih (2008a). e Lim & Teh (2007). First Prev Page 32 Next Last Go Back Full Screen Close Quit

33 Results on Netflix TABLE: Performance on Netflix Data Method RMSE Run Time (hours) Cinematch SVD PMF RBM PMF-VB BPMF BSRM NREM NREM First Prev Page 33 Next Last Go Back Full Screen Close Quit

34 Predictive Uncertainty Standard deviations of prediction residuals vs. standard deviations predicted by our model on EachMovie First Prev Page 34 Next Last Go Back Full Screen Close Quit

35 Related work Multi-task learning using Gaussian processes, those learn the covariance Σ shared across tasks a, and those that additionally consider the covariance Ω between tasks b Application of GP models to collaborative filtering c Low-rank matrix factorization, e.g., d. Our model is nonparametric in the sense no rank constraint is imposed. Very few matrix factorization methods use known predictors. One such a work e introduces low-rank multiplicative random effects in modeling networked observations. a Lawrence & Platt (2004); Schwaighofer, Tresp & Yu (2004); Yu, Tresp & Schwaighofer (2005). b Yu, Chu, Yu, Tresp, & Xu, (2007); Bonilla, Chai, & Williams (2008). c Schwaighofer, Tresp & Yu (2004), Yu & Chu (2007) d Salakhutdinov & Mnih (2008b). e Hoff (2005) First Prev Page 35 Next Last Go Back Full Screen Close Quit

36 Summary The model provides a novel way to use random effects and known attributes to explain the complex dependence of data; We make the nonparametric model scalable and efficient on very large-scale problems; Our experiments demonstrate that the algorithm works very well on two challenging collaborative prediction problems; In the near future, it will be promising to perform a full Bayesian inference by a parallel Gibbs sampling method. First Prev Page 36 Next Last Go Back Full Screen Close Quit

Large-scale Collaborative Prediction Using a Nonparametric Random Effects Model

Large-scale Collaborative Prediction Using a Nonparametric Random Effects Model Kai Yu John Lafferty Shenghuo Zhu Yihong Gong NEC Laboratories America, Cupertino, CA 95014 USA Carnegie Mellon University,