Machine Learning. Principal Components Analysis. Le Song. CSE6740/CS7641/ISYE6740, Fall 2012

Transcription

1 Machine Learning CSE6740/CS7641/ISYE6740, Fall 2012 Principal Components Analysis Le Song Lecture 22, Nov 13, 2012 Based on slides from Eric Xing, CMU Reading: Chap 12.1, CB book 1

2 2 Factor or Component Analysis We study phenomena that can not be directly observed ego, personality, intelligence in psychology Underlying factors that govern the observed data We want to identify and operate with underlying latent factors rather than the observed data E.g. topics in news articles Transcription factors in genomics We want to discover and exploit hidden relationships beautiful car and gorgeous automobile are closely related So are driver and automobile But does your search engine know this? Reduces noise and error in results

3 3 Factor or Component Analysis, cond. We have too many observations and dimensions To reason about or obtain insights from To visualize Too much noise in the data Need to reduce them to a smaller set of factors Better representation of data without losing much information Can build more effective data analyses on the reduced-dimensional space: classification, clustering, pattern recognition Combinations of observed variables may be more effective bases for insights, even if physical meaning is obscure

4 4 The goal: Discover a new set of factors/dimensions/axes against which to represent, describe or evaluate the data For more effective reasoning, insights, or better visualization Reduce noise in the data Typically a smaller set of factors: dimension reduction Better representation of data without losing much information Can build more effective data analyses on the reduced-dimensional space: classification, clustering, pattern recognition Factors are combinations of observed variables May be more effective bases for insights, even if physical meaning is obscure Observed data are described in terms of these factors rather than in terms of original variables/dimensions

5 5 Basic Concept Areas of variance in data are where items can be best discriminated and key underlying phenomena observed Areas of greatest signal in the data If two items or dimensions are highly correlated or dependent They are likely to represent highly related phenomena If they tell us about the same underlying variance in the data, combining them to form a single measure is reasonable Parsimony Reduction in Error So we want to combine related variables, and focus on uncorrelated or independent ones, especially those along which the observations have high variance We want a smaller set of variables that explain most of the variance in the original data, in more compact and insightful form

6 6 Basic Concept What if the dependences and correlations are not so strong or direct? And suppose you have 3 variables, or 4, or 5, or 10000? Look for the phenomena underlying the observed covariance/co-dependence in a set of variables Once again, phenomena that are uncorrelated or independent, and especially those along which the data show high variance These phenomena are called factors or principal components or independent components, depending on the methods used Factor analysis: based on variance/covariance/correlation Independent Component Analysis: based on independence

7 An example: 7

8 Original Variable B Principal Component Analysis Most common form of factor analysis The new variables/dimensions Are linear combinations of the original ones Are uncorrelated with one another Orthogonal in original dimension space Capture as much of the original variance in the data as possible Are called Principal Components PC 2 Original Variable A Orthogonal directions of greatest variance in data Projections along PC1 discriminate the data most along any one axis PC 1 8

9 Original Variable B 9 Principal Component Analysis PC 2 PC 1 First principal component is the direction of greatest variability (covariance) in the data Second is the next orthogonal (uncorrelated) direction of greatest variability So first remove all the variability along the first component, and then find the next direction of greatest variability Original Variable A And so on

10 10 Computing the Components Data points are vectors in a multidimensional space Projection of vector x onto an axis (dimension) u is u T x Direction of greatest variability is that in which the average square of the projection is greatest I.e. u such that E((u T x) 2 ) over all x is maximized (we subtract the mean along each dimension, and center the original axis system at the centroid of all data points, for simplicity) This direction of u is the direction of the first Principal Component

11 Computing the Components E(S i (u T x i ) 2 ) = E ((u T X) (u T X) T ) = E (u T XX T u) The matrix C = XX T contains the correlations (similarities) of the original axes based on how the data values project onto them So we are looking for u that maximizes u T Cu, subject to u being unit-length It is maximized when u is the principal eigenvector of the matrix C, in which case u T Cu = u T lu = l if u is unit-length, where l is the principal eigenvalue of the correlation matrix C The eigenvalue denotes the amount of variability captured along that dimension 11

12 12 Why the Eigenvectors? Maximise u T XX T u s.t u T u = 1 Construct Langrangian u T XX T u λu T u Vector of partial derivatives set to zero xx T u λu = (xx T λi) u = 0 As u 0 then u must be an eigenvector of XX T with eigenvalue λ

13 Computing the Components Similarly for the next axis, etc. So, the new axes are the eigenvectors of the matrix of correlations of the original variables, which captures the similarities of the original variables based on how data samples project to them Geometrically: centering followed by rotation Linear transformation 13

14 14 PCs, Variance and Least-Squares The first PC retains the greatest amount of variation in the sample The k th PC retains the kth greatest fraction of the variation in the sample The k th largest eigenvalue of the correlation matrix C is the variance in the sample along the k th PC The least-squares view: PCs are a series of linear least squares fits to a sample, each orthogonal to all previous ones

15 How Many PCs? For n original dimensions, correlation matrix is nxn, and has up to n eigenvectors. So n PCs. Where does dimensionality reduction come from? Can ignore the components of lesser significance Variance (%) PC1 PC2 PC3 PC4 PC5 PC6 PC7 PC8 PC9 PC10 You do lose some information, but if the eigenvalues are small, you don t lose much n dimensions in original data calculate n eigenvectors and eigenvalues choose only the first p eigenvectors, based on their eigenvalues final data set has only p dimensions 15

16 16 Application: Latent Semantic Analysis Motivation Lexical matching at term level inaccurate (claimed) Polysemy words with number of meanings term matching returns irrelevant documents impacts precision Synonomy number of words with same meaning term matching misses relevant documents impacts recall LSA assumes that there exists a LATENT structure in word usage obscured by variability in word choice Analogous to signal + additive noise model in signal processing

17 17 The Vector Space Model Represent each document by a high-dimensional vector in the space of words

18 18 The Corpora Matrix Doc 1 Doc 2 Doc 3... Word X = Word Word Word Word

19 Feature Vector Representation 19

20 Problems Looks for literal term matches Terms in queries (esp short ones) don t always capture user s information need well Problems: Synonymy: other words with the same meaning Car and automobile No associations between words are made in the vector space representation. Polysemy: the same word having other meanings Apple (fruit and company) The vector space model is unable to discriminate between different meanings of the same word. What if we could match against concepts, that represent related words, rather than words themselves 20

21 21 Example of Problems -- Relevant docs may not have the query terms but may have many related terms -- Irrelevant docs may have the query terms but may not have any related terms

22 22 Latent Semantic Indexing (LSI) Uses statistically derived conceptual indices instead of individual words for retrieval Assumes that there is some underlying or latent structure in word usage that is obscured by variability in word choice Key idea: instead of representing documents and queries as vectors in a t-dim space of terms Represent them (and terms themselves) as vectors in a lower-dimensional space whose axes are concepts that effectively group together similar words Uses SVD to reduce document representations, The axes are the Principal Components from SVD So what is SVD? (Deerwester et al., 1990)

23 23 Example Suppose we have keywords Car, automobile, driver, elephant We want queries on car to also get docs about drivers and automobiles, but not about elephants What if we could discover that the cars, automobiles and drivers axes are strongly correlated, but elephants is not How? Via correlations observed through documents If docs A & B don t share any words with each other, but both share lots of words with doc C, then A & B will be considered similar E.g A has cars and drivers, B has automobiles and drivers When you scrunch down dimensions, small differences (noise) gets glossed over, and you get desired behavior

24 Term 24 Latent Semantic Indexing Document =... * * X (m x n) T (m x k) w = K k=1 d k l T L (k x k) k k D T (k x n) This is our compressed representation of a document

25 Eigenvalues & Eigenvectors Eigenvectors (for a square m m matrix S) Example (right) eigenvector eigenvalue How many eigenvalues are there at most? only has a non-zero solution if this is a m-th order equation in λ which can have at most m distinct solutions (roots of the characteristic polynomial) can be complex even though S is real. 25

26 26 Eigenvalues & Eigenvectors For symmetric matrices, eigenvectors for distinct eigenvalues are orthogonal Sv { 1, 2} l { 1, 2} v{ 1, 2 and l1 l2 v1 v2 }, All eigenvalues of a real symmetric matrix are real. 0 for complex l, if S li 0 and S T S l All eigenvalues of a positive semidefinite matrix are nonnegative w n T, w Sw 0, then if Sv lv l 0

27 27 Eigen/diagonal Decomposition Let be a square matrix with m linearly independent eigenvectors (a non-defective matrix) Theorem: Exists an eigen decomposition diagonal Unique for distinct eigenvalues (cf. matrix diagonalization theorem) Columns of U are eigenvectors of S Diagonal elements of are eigenvalues of

28 Singular Value Decomposition For an m n matrix A of rank r there exists a factorization (Singular Value Decomposition = SVD) as follows: A USV T m m m n V is n n The columns of U are orthogonal eigenvectors of AA T. The columns of V are orthogonal eigenvectors of A T A. Eigenvalues l 1 l r of AA T are the eigenvalues of A T A. i l i S diag 1... r Singular values. 28

29 29 SVD and PCA The first root is called the prinicipal eigenvalue which has an associated orthonormal (u T u = 1) eigenvector u Subsequent roots are ordered such that λ 1 > λ 2 > > λ M with rank(d) non-zero values. Eigenvectors form an orthonormal basis i.e. u it u j = δ ij The eigenvalue decomposition of XX T = UΣU T where U = [u 1, u 2,, u M ] and Σ = diag[λ 1, λ 2,, λ M ] Similarly the eigenvalue decomposition of X T X = VΣV T The SVD is closely related to the above X=U Σ 1/2 V T The left singular vectors U, right singular vectors V, singular values = square root of eigenvalues.

30 Example term ch2 ch3 ch4 ch5 ch6 ch7 ch8 ch9 controllability observability realization feedback controller observer transfer function polynomial matrices This happens to be a rank-7 matrix -so only 7 dimensions required U (9x7) = S (7x7) = Singular values = Sqrt of Eigen values of AA T V (7x8) = T

31 Low-rank Approximation Solution via SVD A Udiag(,...,,0,..., 0) V k 1 k T set smallest r-k singular values to zero k A k k i 1 i uv i T i column notation: sum of rank 1 matrices 31

32 32 Approximation error How good (bad) is this approximation? It s the best possible, measured by the Frobenius norm of the error: X rank( X) k : min A X F A A k F k 1... m where the i are ordered such that i i+1. Suggests why Frobenius error drops as k increased.

33 Term 33 SVD Low-rank approximation Whereas the term-doc matrix A may have m=50000, n=10 million (and rank close to 50000) We can construct an approximation A 100 with rank 100. Of all rank 100 matrices, it would have the lowest Frobenius error. Document =... * * X (m x n) T (m x k) L (k x k) D T (k x n) w = C. Eckart, G. Young, The approximation of a matrix by another of lower rank. Psychometrika, 1, , K k=1 d k l T k k

34 Following the Example term ch2 ch3 ch4 ch5 ch6 ch7 ch8 ch9 controllability observability realization feedback controller observer transfer function polynomial matrices This happens to be a rank-7 matrix -so only 7 dimensions required U (9x7) = S (7x7) = Singular values = Sqrt of Eigen values of AA T V (7x8) = T

35 PCs can be viewed as Topics In the sense of having to find quantities that are not observable directly Similarly, transcription factors in biology, as unobservable causal bridges between experimental conditions and gene expression 35

36 36 Querying To query for feedback controller, the query vector would be q = [ ]' (' indicates transpose), Let q be the query vector. Then the document-space vector corresponding to q is given by: q'*u2*inv(s2) = Dq Point at the centroid of the query terms poisitions in the new space. For the feedback controller query vector, the result is: Dq = To find the best document match, we compare the Dq vector against all the document vectors in the 2- dimensional V2 space. The document vector that is nearest in direction to Dq is the best match. The cosine values for the eight document vectors and the query vector are: U2 (9x2) = S2 (2x2) = V2 (8x2) = term ch2 ch3 ch4 ch5 ch6 ch7 ch8 ch9 controllability observability realization feedback controller observer transfer function polynomial matrices

37 Medline data 37

38 K is the number of singular values used Within.40 threshold 38

39 What LSI can do LSI analysis effectively does Dimensionality reduction Noise reduction Exploitation of redundant data Correlation analysis and Query expansion (with related words) Some of the individual effects can be achieved with simpler techniques (e.g. thesaurus construction). LSI does them together. LSI handles synonymy well, not so much polysemy Challenge: SVD is complex to compute (O(n 3 )) Needs to be updated as new documents are found/updated 39

40 40 Summary: Principle Linear projection method to reduce the number of parameters Transfer a set of correlated variables into a new set of uncorrelated variables Map the data into a space of lower dimensionality Form of unsupervised learning Properties It can be viewed as a rotation of the existing axes to new positions in the space defined by original variables New axes are orthogonal and represent the directions with maximum variability Application: In many settings in pattern recognition and retrieval, we have a feature-object matrix. For text, the terms are features and the docs are objects. Could be opinions and users This matrix may be redundant in dimensionality. Can work with low-rank approximation. If entries are missing (e.g., users opinions), can recover if dimensionality is low.