Probabilistic Latent Semantic Analysis

Size: px

Start display at page:

Download "Probabilistic Latent Semantic Analysis"

Maude Henry
6 years ago
Views:

1 Probabilistic Latent Semantic Analysis Seungjin Choi Department of Computer Science and Engineering Pohang University of Science and Technology 77 Cheongam-ro, Nam-gu, Pohang 37673, Korea 1 / 27

2 Outline Singular value decomposition Latent semantic analysis (a.k.a. latent semantic indexing) Probabilistic latent semantic analysis (a.k.a. probabilistic latent semantic indexing) 2 / 27

3 Range Space and Null Space Definition For any A R m n, its range space R(A) and null space N (A) are defined as follows: R(A) = {y R m y = Ax for some x R n } N (A) = {x R n Ax = 0}. Examples: What are range space and null space of A and B? [ ] [ ] A = B = / 27

4 Linear Algebraic Equations A linear algebraic equation has the form where A R m n, x R n, and b R m. Ax = b, (1) Theorem A solution of Ax = b exists if and only if b R(A). Theorem Let x be a particular solution to (1), then x = x + N (A) = {x + w w N (A)} is the general solution. Remark: Note that N (A) is the set of homogeneous solutions to (1). Therefore the set of all solutions is the sum of a particular solution and the set of homogeneous solutions. 4 / 27

5 Singular Value Decomposition Theorem (SVD) Let A R m n. Then there exist orthogonal matrices such that U = [u 1, u 2,..., u m ] R m m V = [v 1, v 2,..., v n ] R n n U T AV = diag (σ 1, σ 2,..., σ p ) R m n (2) where p = min{m, n} and σ 1 σ 2 σ p 0. The scalars {σ i R} are called singular values. SVD A = UΣV 5 / 27

6 Low-Rank Approximation Theorem (Eckart and Young, 1936) Suppose that A is an m n matrix of rank r, with singular values σ 1 σ 2 σ r. and SVD A = UΣV. Then the best rank-k approximation of A is A k = U 1 Σ 1 V 1, where That is, A A k F = min U = [U 1, U 2 ], V = [V 1, V 2 ], Σ 1 = diag(σ 1,..., σ k ). { } A Â F Â Rm n, rank(â) = k. 6 / 27

7 Proof of Eckart-Young Theorem Note that Direct calculations give A Â F = UΣV Â F = Σ U } {{ ÂV } F. N Σ N 2 F = i,j Σ i,j N i,j 2 = r σ i N i,i 2 + N i,i 2 + i>r i j i=1 N i,j 2, which is minimal when all the non-diagonal entries of N are equal to zero and so are all N i,i for i > r. The minimum of r i=1 σ i N i,i 2 is attained when σ i = N i,i for i = 1,..., k and all other N i,i are zero. 7 / 27

8 Least Squares Problem Consider a matrix A R m n with Rank(A) = r. Then R(A) = R(U 1 ) = sp [u 1, u 2,..., u r ] N (A) = sp [v r+1, v r+2,..., v n ] = R(V 2 ). The least squares (LS) problem is as follows: Setting x b Ax 2 = 0 leads to min x R n b Ax 2. (3) A T Ax = A T b. (4) 8 / 27

9 Remarks From SVD, one can see that the particular solution to (4) is x = V 1 Σ 1 U T 1 b. The complete solution is x = x + N (A T A). Note that N (A T A) = N (A) = sp[v 2 ]. Hence x = V 1 Σ 1 U T 1 b + V 2 w for some w R r. Since V 1 Σ 1 U T 1 b V 2 w for all w, x = V 1 Σ 1 U T 1 b is the MMSE solution for which x is the smallest among all solutions. The pseudo-inverse of A is The condition number of A is given by σ1 σ r. A = V 1 Σ 1 U T 1. (5) 9 / 27

10 PCA Principal component analysis (PCA) is a well-established technique for dimension reduction. Its applications include data compression, image processing, data visualization, exploratory data analysis, pattern recognition, and time series prediction. The most common derivation of PCA is in terms of a orthogonal projection which maximizes the variance in the projected space. Given a set of m-dimensional observation vector, {x t }, the PCA aims at finding a orthogonal linear projection y = W x such that the variance of y R q (q < d) is maximized. The ith element of y is called the ith principal component. Alternatively, the PCA provides an orthogonal linear projection which minimize the squared reconstruction error t x t ˆx t 2. Thus the PCA is an optimal linear encoding in MS sense. 10 / 27

11 PCA and SVD It was shown that the ith row vector of W denoted by w T i corresponds to the normalized eigenvector associated with the ith largest eigenvalue of the covariance matrix R x = E{xx T } Principal components can be found by SVD, linear neural networks, or probabilistic methods. The SVD of R x has the form R x = UΣV T. Then we select n column vectors to contruct U 1 = [u 1,..., u n ]. The the PCA transform leads to y = U T 1 x. 11 / 27

12 PCA: An Example / 27

13 Spectral Decomposition (Eigen-Decomposition) Given a symmetric matrix C R m m, its spectral decomposition is given by C = λ 1 u 1 u T 1 + λ 2 u 2 u T λ m u m u T m, where λ 1 λ 2 λ m are eigenvalues of C and u i are associated eigenvectors. 13 / 27

14 Power Iteration The power iteration is a classical method which finds the largest eigenvector (associated with the largest eigenvalue) of a matrix C R m m. Given a symmetric matrix C R m m (hence its eigenvalues are real), the power iteration starts from a nonzero vector w(0) and iteratively updates w(t) by where 2 represents Euclidean norm. w(t + 1) = Cw(t), (6) w(t + 1) = w(t + 1), w(t + 1) 2 (7) Combining (6) and (7) leads to the updating rule which has the form [ w(t + 1) = Cw(t) w T (t)c w(t)] (8) Assume that C has an unique eigenvalue of maximum modulus λ 1 associated with the leading eigenvector u 1. Then the power iteration (8) leads w(t) to converge to u / 27

15 Deflation Suppose that we are interested in computing eigenvectors of the data covariance matrix C = xx T. The power iteration is applied to C for extracting its first eigenvector. Our question arises, How can we compute the second eigenvector of C using the power iteration?. The deflation method is a common numerical technique for computing several eigenvalues and eigenvectors of C. Assume that the first eigenvector is already computed. Then the output value can be deflated by the following transformation: x = ( I u 1 u T ) 1 x. (9) One can easily see that x x T = m i=2 λ iu i u T i. The power iteration is applied to the deflated data, in order to extract the second eigenvector of C. 15 / 27

16 Term-Document Matrix A term-document matrix X R D N is a collection of vector space representations of documents, where rows are terms (words) and columns are documents ( ) N X ij = t ij log, idf i where t ij is the term frequency of word i in document j and idf i is the number of documents containing word i. We write X = [d 1, d 2,..., d N ] (document vectors) = t 1 t 2. t D (term vectors). 16 / 27

17 Latent Semantic Analysis A method for automatic indexing and retrieval, uncovering latent semantic structure of a term-document matrix. SVD of X R D N is given by X UΣV where U R D K, Σ R K K, and V R N K. D N K K X U Σ V K K D N Comparing documents: Σ 1 U X = Σ 1 U UΣV = V = [ d 1,..., d N ]. Comparing terms: X V Σ 1 = UΣV V Σ 1 = U = [ t 1,..., t D ]. 17 / 27

18 Probabilistic Latent Semantic Analysis The key idea in latent semantic analysis is to map high-dimensional count vectors (co-occurrence data, dyadic data) to a lower-dimensional representation in a so-called latent semantic space. The goal of LSA is to find a data mapping which reveals semantical relations between the entries of interest. Probabilistic latent semantic analysis (PLSA) is a probabilistic variant of LSA: Has a sound statistical foundation; Defines a proper generative model of the data. Aspect model: A latent variable model for co-occurrence data which associates an unobserved class z {z 1,..., z K } with each occurrence of a word w {w 1,..., w D } in a document d {d 1,..., d N }. 18 / 27

19 PLSA: Graphical Representation d j A document d j and a term (word) w i are conditionally independent given an unobserved topic z: [ ] z p(w i, d j ) = p(d j ) Generation process: z p(w i z)p(z d j ). w i N D Select a document dj with probability p(d j ). Pick a latent class (topic) with probability p(z d j ). Generate a word wi with probability p(w i z). 19 / 27

20 PLSA: Symmetric Parameterization d j z p(w i, d j ) = z = z p(w i, d j, z) p(w i, d j z)p(z) N = z p(w i z)p(d j z)p(z). w i D 20 / 27

21 Model Fitting: EM Algorithm Dyadic data X : Entries X ij are made for dyads (w i, d j ) which refer to a domain with two sets of objects, W = {w 1,..., w D } and D = {d 1,..., d N }. Complete-data likelihood: p(x, z) = p(w i, d j, z) C ij = [p(w i z)p(d j z)p(z)] C ij, i j where C ij are the empirical counts for dyads (w i, d j ) and X ij = C ij / i j C ij. EM optimization E-step: Compute the expected complete-data log-likelihood: L c = p(z k w i, d j )C ij log [p(w i z k )p(d j z k )p(z k )]. i j k M-step: Re-estimate parameters p(wi z k ), p(d j z k ), p(z k ) which maximizes L c. 21 / 27

22 E-Step: Compute p(z w i, d j ) Compute the posterior distribution over latent variables: p(z k w i, d j ) = p(w i, d j z k )p(z k ) p(w i, d j ) p(w i z k )p(d j z k )p(z k ) = l p(w i z l )p(d j z l )p(z l ), where p(w i z k ), p(d j z k ), p(z k ) are estimated in the M-step. 22 / 27

23 M-Step: Re-estimate Parameters Re-estimate parameters: p(w i z k ) = p(d j z k ) = p(z k ) = j C ijp(z k w i, d j ) j C ijp(z k w i, d j ), i i C ijp(z k w i, d j ) j C ijp(z k w i, d j ), i i j C ijp(z k w i, d j ) i j C. ij 23 / 27

24 Updating equations in the M-step are determined by solving [ L c + λ(1 ] p(w i z k )) = 0, p(w i z k ) i L c + λ(1 p(d j z k )) = 0, p(d j z k ) j [ L c + λ(1 ] p(z l )) = 0, p(z k ) l for p(w i z k ), p(d j z k ), p(z k ), respectively. 24 / 27

25 Document Clustering by PLSA You are given parameters p(w i z k ), p(d j z k ), p(z k ) estimated by EM optimization. Compute p(z k d j ) p(d j z k )p(z k ). Assign document d j to cluster k if k = arg max p(z k d j ). k 25 / 27

26 PLSA: Revisited The PLSA models each word in a document as a sample from a mixture model where the mixture components are multinomial random variables that can be viewed as representations of topics, p(w, d) = z p(z)p(w z)p(d z) = p(d) z p(w z)p(z d). Each document is represented as a list of mixing proportions for mixture components and reduced to a probability distribution on a fixed set of topics. This distribution is the reduced description associated with the document. In PLSA, each document is represented as a list of numbers (the mixing proportions for topics) and there is no generative probabilistic model for these numbers. This leads to the following problems: The number of parameters in the model grows linearly with the size of corpus, which leads to problems with overfitting. It is not clear how to assign probability to a document outside of the training set. 26 / 27

27 References S. Deerwester, S. T. Dumais, and R. Harshman, Indexing by latent semantic analysis, Journal of the Americal Society of Information Science, vol. 41, no. 6, pp , T. Hofmann, Probabilistic latent semantic indexing, in Proc. SIGIR T. Hofmann, Probabilistic latent semantic analysis, in Proc. UAI / 27

Notes on Latent Semantic Analysis

Notes on Latent Semantic Analysis Costas Boulis 1 Introduction One of the most fundamental problems of information retrieval (IR) is to find all documents (and nothing but those) that are semantically