Introduction to Kernel Principal Component Analysis(PCA) Mohammed Nasser Dept. of Statistics, RU,Bangladesh Email: mnasser.ru@gmail.com 1 Contents Basics of PCA Application of PCA in Face Recognition Some Terms in PCA Motivation for KPCA Basics of KPCA Applications of KPCA High-dimensional Data Gene expression Face images Handwritten digits Why Feature Reduction? • Most machine learning and data mining techniques may not be effective for high-dimensional data – Curse of Dimensionality – Query accuracy and efficiency degrade rapidly as the dimension increases. • The intrinsic dimension may be small. – For example, the number of genes responsible for a certain type of disease may be small. Why Reduce Dimensionality? 1. 2. 3. 4. 5. 6. Reduces time complexity: Less computation Reduces space complexity: Less parameters Saves the cost of observing the feature Simpler models are more robust on small datasets More interpretable; simpler explanation Data visualization (structure, groups, outliers, etc) if plotted in 2 or 3 dimensions Feature reduction algorithms • Unsupervised – Latent Semantic Indexing (LSI): truncated SVD – Independent Component Analysis (ICA) – Principal Component Analysis (PCA) – Canonical Correlation Analysis (CCA) • Supervised – Linear Discriminant Analysis (LDA) • Semi-supervised – Research topic Algebraic derivation of PCs • Main steps for computing PCs – Form the covariance matrix S. – Compute its eigenvectors: u i i 1 – Use the first d eigenvectors to form the d PCs. p u i i 1 d – The transformation G is given by G [ u1 , u 2 , A test point x G x . p T d , ud ] Optimality property of PCA Reconstruction Dimension reduction X p n G T X d n Original data G T G X T d n X G (G X ) T pn d p Y G X T X X pn pn G pd d n Optimality property of PCA Main theoretical result: The matrix G consisting of the first d eigenvectors of the covariance matrix S solves the following min problem: min X G (G X ) T G pd 2 subject to F X X G G Id T 2 F reconstruction error PCA projection minimizes the reconstruction error among all linear projections of size d. Dimensionality Reduction • One approach to deal with high dimensional data is by reducing their dimensionality. • Project high dimensional data onto a lower dimensional sub-space using linear or non-linear transformations. Dimensionality Reduction • Linear transformations are simple to compute and tractable. t Y U X ( bi u i a i ) kx1 dx1 kxd (k<<d) • Classical –linear- approaches: – Principal Component Analysis (PCA) – Fisher Discriminant Analysis (FDA) –Singular Value Decomosition (SVD) --Factor Analysis (FA) --Canonical Correlation(CCA) Principal Component Analysis (PCA) • Each dimensionality reduction technique finds an appropriate transformation by satisfying certain criteria (e.g., information loss, data discrimination, etc.) • The goal of PCA is to reduce the dimensionality of the data while retaining as much as possible of the variation present in the dataset. Principal Component Analysis (PCA) • Find a basis in a low dimensional sub-space: – Approximate vectors by projecting them in a low dimensional sub-space: (1) Original space representation: x a1 v1 a 2 v 2 ... a N v N w h ere v1 , v 2 , ..., v n is a b ase in th e o rig in al N -d im en si o n al sp ace (2) Lower-dimensional sub-space representation: xˆ b1u 1 b 2 u 2 ... b K u K w h ere u 1 , u 2 , ..., u K is a b ase in th e K -d im en sio n al su b -s p ace (K < N ) • Note: if K=N, then xˆ x Principal Component Analysis (PCA) • Example (K=N): Principal Component Analysis (PCA) • Methodology – Suppose x1, x2, ..., xM are N x 1 vectors Principal Component Analysis (PCA) • Methodology – cont. bi u i ( x x ) T Principal Component Analysis (PCA) • Linear transformation implied by PCA – The linear transformation RN RK that performs the dimensionality reduction is: Principal Component Analysis (PCA) • How many principal components to keep? – To choose K, you can use the following criterion: Unfortunately for some data sets to meet this requirement we need K almost equal to N. That is, no effective data reduction is possible. Principal Component Analysis (PCA) • Eigenvalue spectrum K λ Scree iplot λN Principal Component Analysis (PCA) • Standardization – The principal components are dependent on the units used to measure the original variables as well as on the range of values they assume. – We should always standardize the data prior to using PCA. – A common standardization method is to transform all the data to have zero mean and unit standard deviation: CS 479/679 Pattern Recognition – Spring 2006 Dimensionality Reduction Using PCA/LDA Chapter 3 (Duda et al.) – Section 3.8 Case Studies: Face Recognition Using Dimensionality Reduction M. Turk, A. Pentland, "Eigenfaces for Recognition", Journal of Cognitive Neuroscience, 3(1), pp. 71-86, 1991. D. Swets, J. Weng, "Using Discriminant Eigenfeatures for Image Retrieval", IEEE Transactions on Pattern Analysis and Machine Intelligence, 18(8), pp. 831-836, 1996. A. Martinez, A. Kak, "PCA versus LDA", IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 23, no. 2, pp. 228-233, 2001. Principal Component Analysis (PCA) • Face Recognition – The simplest approach is to think of it as a template matching problem – Problems arise when performing recognition in a high-dimensional space. – Significant improvements can be achieved by first mapping the data into a lower dimensionality space. – How to find this lower-dimensional space? Principal Component Analysis (PCA) • Main idea behind eigenfaces average face Principal Component Analysis (PCA) • Computation of the eigenfaces Principal Component Analysis (PCA) • Computation of the eigenfaces – cont. Principal Component Analysis (PCA) • Computation of the eigenfaces – cont. Mind that this is normalized.. ui Principal Component Analysis (PCA) • Computation of the eigenfaces – cont. Principal Component Analysis (PCA) • Representing faces onto this basis Principal Component Analysis (PCA) • Representing faces onto this basis – cont. Principal Component Analysis (PCA) • Face Recognition Using Eigenfaces Principal Component Analysis (PCA) • Face Recognition Using Eigenfaces – cont. – The distance er is called distance within the face space (difs) – Comment: we can use the common Euclidean distance to compute er, however, it has been reported that the Mahalanobis distance performs better: Principal Component Analysis (PCA) • Face Detection Using Eigenfaces Principal Component Analysis (PCA) • Face Detection Using Eigenfaces – cont. Principal Components Analysis So, principal components are given by: b1 = u11x1 + u12x2 + ... + u1NxN b2 = u21x1 + u22x2 + ... + u2NxN ... bN= aN1x1 + aN2x2 + ... + aNNxN xj’s are standardized if correlation matrix is used (mean 0.0, SD 1.0) Score of ith unit on jth principal component bi,j = uj1xi1 + uj2xi2 + ... + ujNxiN PCA Scores 5 xi2 bi,1 4 bi,2 3 2 4.0 4.5 5.0 xi1 5.5 6.0 Principal Components Analysis Amount of variance accounted for by: 1st principal component, λ1, 1st eigenvalue 2nd principal component, λ2, 2ndeigenvalue ... λ1 > λ2 > λ3 > λ4 > ... Average λj = 1 (correlation matrix) Principal Components Analysis: Eigenvalues 5 λ2 λ1 4 3 U1 2 4.0 4.5 5.0 5.5 6.0 PCA: Terminology • jth principal component is jth eigenvector of correlation/covariance matrix • coefficients, ujk, are elements of eigenvectors and relate original variables (standardized if using correlation matrix) to components • scores are values of units on components (produced using coefficients) • amount of variance accounted for by component is given by eigenvalue, λj • proportion of variance accounted for by component is given by λj / Σ λj • loading of kth original variable on jth component is given by ujk √λj --correlation between variable and component Principal Components Analysis • Covariance Matrix: – Variables must be in same units – Emphasizes variables with most variance – Mean eigenvalue ≠1.0 – Useful in morphometrics, a few other cases • Correlation Matrix: – Variables are standardized (mean 0.0, SD 1.0) – Variables can be in different units – All variables have same impact on analysis – Mean eigenvalue = 1.0 PCA: Potential Problems • Lack of Independence – NO PROBLEM • Lack of Normality – Normality desirable but not essential • Lack of Precision – Precision desirable but not essential • Many Zeroes in Data Matrix – Problem (use Correspondence Analysis) Principal Component Analysis (PCA) • PCA and classification (cont’d) 0 -2 -4 v 2 4 Motivation -3 -2 -1 0 z 1 2 3 Motivation 0 2 4 u 6 8 ??????? -3 -2 -1 0 z 1 2 3 Motivation Linear projections will not detect the pattern. Limitations of linear PCA 1,2,3=1/3 Nonlinear PCA Three popular methods are available: 1) Neural-network based PCA (E. Oja, 1982) 2)Method of Principal Curves (T.J. Hastie and W. Stuetzle, 1989) 3) Kernel based PCA (B. Schölkopf, A. Smola, and K. Müller, 1998) PCA NPCA Kernel PCA: The main idea A Useful Theorem for Hilbert space Let H be a Hilbert space and x1, ……xn in H. Let M=span{x1, ……xn}. Also u and v in M. <xi,u>=<xi,v>, i=1,……,n implies u=v Proof. Try your self. Kernel methods in PCA Linear PCA Cw w ( 1) where C is covariance matrix for centered data X: C 1 n Cw 1 l ' xixi i 1 n n (x i ' w ) x i w i 1 w span{ x1 , ..... x 2 } if 0 xi , w xi , Cw i=1......l (2) (1) and (2) are equivalent conditions. Kernel methods in PCA Now let us suppose: :R p F , t h e fea t u r e sp a ce Possibly F is a very high dimension space. In Kernel PCA, we do the PCA in feature space. C 1 l l ( x i ) ( x i ) T ( w h a t is it s m ea n in g? ? ) i 1 v C v 1 l l (xi ), v (xi ) i 1 remember about centering! (*) Kernel Methods in PCA Again all solutions generated by v with { ( x i ) , 0 lie in the space , (x l ) } It has two useful consequences: v sp a n o f { ( x i ) , 1} , (x l ) } l v i (x i ) i 1 2) We may instead solve the set of equations ( xi ), v ( xi ), Cv i=1......l Kernel Methods in PCA Defining an lxl kernel matrix K: k x i , x j ( x i ), ( x j ) And using the result (1) in ( 2) we get lK K 2 ( 3) But we need not solve (3). It can be shown easily that the following simpler system gives us solutions that are interesting to us. l K ( 4) Kernel Methods in PCA Compute eigenvalue problem for the kernel matrix Kα α The solutions (k, k) further need to be normalized by imposing k , k 1 sin ce v k k sh ou ld b e w it h v k 1 If x is our new observation, the feature value (??) will be ( x ) and kth principal score will be l v , ( x ) k i ( x ), ( x i ) k i 1 l i K (x , x i ) k i 1 Kernel Methods in PCA Data centering: ˆ ( x ) ( x ) S ( x ) ( x ) l 1 (x l i ) i 1 Hence, the kernel for the transformed space is kˆ ( x, z ) ˆ ( x ), ˆ ( z ) ( x ) 1 l (x l i ) , ˆ ( z ) 1 i 1 k ( x, z ) 1 l k (x, x l i 1 i ) 1 l k ( z, x l i 1 i ) l (x l i 1 1 l 2 l k (x i , j 1 i ,x j) i ) Kernel Methods in PCA Expressed as an operation on the kernel matrix this can be rewritten as 1 1 1 ˆ K K j j' K K j j' 2 (j' K j) j j' l l l where j is the all 1s vector. Linear PCA Kernel PCA captures the nonlinear structure of the data Linear PCA Kernel PCA captures the nonlinear structure of the data Algorithm Input: Data X={x1, x2, …, xl} in n-dimensional space. Process: Ki,j= k(xi,xj); i,j=1,…, l. ˆ K 1 j j' K 1 K j j' 1 ( j' K j) j j'; K 2 l l l ˆ ); [V , ] eig( K (j ) 1 j vj, … for centered data j 1, ..., l . k (j ) x j i k(xi, x) i 1 j 1 l Kernel matrix ... Output: Transformed data k-dimensional vector projection of new data into this subspace Reference • I.T. Jolliffe. (2002)Principal Component Analysis. • . Schölkopf, et al. (1998 Kernel Principal Component Analysis)/ • B. . Schölkopf and A.J. Smola(2000/20012002) Learning with Kernels • Christopher J C Burges (2005).Geometric Methods for Feature Extraction and Dimensional Reduction.