Robust estimation of principal components from depth-based multivariate rank covariance matrix

Transcription

1 Robust estimation of principal components from depth-based multivariate rank covariance matrix Subho Majumdar Snigdhansu Chatterjee University of Minnesota, School of Statistics

2 Table of contents

3 Summary Introduction: what is data depth? Multivariate ranks based on data depth The Depth Covariance Matrix (DCM): overview of results Performance: simulations and real data analysis

4 What is depth? Example: 500 points from N 2 ((0, 0) T, diag(2, 1)) A scalar measure of how much inside a point is with respect to a data cloud

5 Formal definition For any multivariate distribution F = F X, the depth of a point x R p, say D(x, F X ) is any real-valued function that provides a center outward ordering of x with respect to F (Zuo and Serfling, 2000). Desirable properties (Liu, 1990) (P1) Affine invariance: D(Ax + b, F AX+b ) = D(x, F X ) (P2) Maximality at center: D(θ, F X ) = sup x R p D(x, F X ) for F X with center of symmetry θ, the deepest point of F X. (P3) Monotonicity w.r.t. deepest point: D(x; F X ) D(θ + a(x θ), F X ) (P4) Vanishing at infinity: D(x; F X ) 0 as x.

6 Examples Halfspace depth (HD) (Tukey, 1975) is the minimum probability of all halfspaces containing a point. HD(x, F ) = inf u R p ;u 0 P(uT X u T x) Projection depth (PD) (Zuo, 2003) is based on an outlyingness function: O(x, F ) = sup u =1 u T x m(u T X) s(u T ; PD(x, F ) = X) O(x, F)

7 Utility of data depth Robustness Classification Depth-weighted means and covariance matrices What we re going to do: Define multivariate rank vectors based on data depth, do PCA on them

8 Spatial signs (Locantore et al., 1999) S(x) = { x x 1 if x 0 0 if x = 0 Say x follows an elliptic distribution with mean µ, covariance matrix Σ. Sign covariance matrix (SCM): Σ S (X) = ES(X µ)s(x µ) T SCM has same eigenvectors as Σ. PCA using SCM is robust, but not efficient X[,1] X[,2] ux[,1] ux[,2]

9 Spatial ranks Fix a depth function D(x, F) = D X (x). Define D X (x) = sup x R p D X (x) D X (x) Transform the original observation: x = D X (x)s(x µ). This is the Spatial Rank of x. Depth Covariance Matrix (DCM) = Cov( X). Has more information than spatial signs, so more efficient X[,1] X[,2] Xrank[,1] Xrank[,2]

10 Form of DCM Theorem (1) Let the random variable X R p follow an elliptical distribution with center µ and covariance matrix Σ = ΓΛΓ T, its spectral decomposition. Then, given a depth function D X (.) the covariance matrix of the transformed random variable X is Cov( X) = ΓΛ D,S Γ T, with Λ D,S = E [( D Z (z)) 2 Λ1/2 zz T Λ 1/2 ] z T (1) Λz where z = (z 1,..., z p ) T N(0, I p ) and Λ D,S a diagonal matrix with diagonal entries ] [( D Z (z)) 2 λ i zi 2 λ D,S,i = E Z p j=1 λ jzj 2 TL; DR: Population eigenvectors are invariant under the spatial rank transformation.

11 Other results Asymptotic distribution of sample DCM, form of its asymptotic variance Asymptotic joint distribution of eigenvectors and eigenvalues of sample DCM Form and shape of influence function: a measure of robustness Asymptotic efficiency relative to sample covariance matrix

12 Simulation 6 elliptical distributions: p-variate normal and t- distributions with df = 5, 6, 10, 15, 25. All distributions centered at 0 p, and have covariance matrix Σ = diag(p, p 1,...1). 3 choices of p: 2, 3 and samples each for sample sizes n = 20, 50, 100, 300, 500 For estimates ˆγ 1 of the first eigenvector ˆγ 1, prediction error is measured by the average smallest angle between the two lines, i.e. Mean Squared Prediction Angle: MSPA(ˆγ 1 ) = m=1 ( cos 1 γ T 1 ˆγ (m) Finite sample efficiency of some eigenvector estimate ˆγ E 1 relative to that obtained from the sample covariance matrix, say ˆγ Cov 1 is: 1 ) FSE(ˆγ E 1, ˆγ Cov 1 ) = MSPA(ˆγCov MSPA(ˆγ E 1 ) 1 ) 2

13 Table of FSE for p = 2 (HSD = Halfspace depth, MhD = Mahalanobis depth, PD = Projection depth) F = Bivariate t 5 SCM HSD-CM MhD-CM PD-CM n= n= n= n= n= F = Bivariate t 6 SCM HSD-CM MhD-CM PD-CM n= n= n= n= n= F = Bivariate t 10 SCM HSD-CM MhD-CM PD-CM n= n= n= n= n=

14 Table of FSE for p = 2 (HSD = Halfspace depth, MhD = Mahalanobis depth, PD = Projection depth) F = Bivariate t 15 SCM HSD-CM MhD-CM PD-CM n= n= n= n= n= F = Bivariate t 25 SCM HSD-CM MhD-CM PD-CM n= n= n= n= n= F = BVN SCM HSD-CM MhD-CM PD-CM n= n= n= n= n=

15 Data analysis: Bus data Features extracted from images of 213 buses: 18 variables Methods compared: Classical PCA (CPCA) SCM PCA (SPCA) ROBPCA (Hubert et al., 2005) PCA based on MCD (MPCA) PCA based on projection-dcm (DPCA)

16 Score distance (SD) and orthogonal distance (OD) SD i = k sij 2 ; λ j j=1 OD i = x i Ps T i where S n k = (s 1,..., s n) T is the scoring matrix, P p k loading matrix, λ 1,..., λ k are eigenvalues obtained from the PCA, and x 1,..., x n are the observation vectors.

17 Bus data: comparison tables Quantile Method of PCA CPCA SPCA ROBPCA MPCA DPCA 10% % % % % % % % % Max Table lists quantiles of the squared orthogonal distance for a sample point from the hyperplane formed by top 3 PCs, For DPCA, more than 90% of points have a smaller orthogonal distance than CPCA

18 Data analysis: Octane data 226 variables and 39 observations. Each observation is a gasoline sample with a certain octane number, and have their NIR absorbance spectra measured in 2 nm intervals between nm. 6 outliers: compounds 25, 26 and 36-39, which contain alcohol. Classical PCA Depth PCA Orthogonal distance Orthogonal distance Score distance Score distance

19 Extensions: Robust kernel PCA 20 points from each person. Noise added to one image from each person. Columns due to kernel CPCA, SPCA and DPCA, respectively. Rows due to top 2, 4, 6, 8 or 10 PCs considered.

20 Extensions: future work Explore properties of a depth-weighted M-estimator of scale matrix: [ ( Σ Dw = E D ] X (x)) 2 (x µ)(x µ) T (x µ) T Σ 1 Dw (x µ) Leverage the idea of depth based ranks: robust non-parametric testing Extending to high-dimensional and functional data

21 Acknowledgements NSF grants # IIS , # SES ; NASA grant # ; The Institute on the Environment (IonE), and College of Liberal Arts (CLA) - University of Minnesota; Abhirup Mallik; My team in Santander Consumer USA;

22 References M. Hubert, P. J. Rousseeuw, and K. V. Branden. ROBPCA: A New Approach to Robust Principal Component Analysis. Technometrics, 47-1:64 79, R.Y. Liu. On a notion of data depth based on random simplices. Ann. of Statist., 18: , N. Locantore, J.S. Marron, D.G. Simpson, N. Tripoli, J.T. Zhang, and K.L. Cohen. Robust principal components of functional data. TEST, 8:1 73, J.W. Tukey. Mathematics and picturing data. In R.D. James, editor, Proceedings of the International Congress on Mathematics, volume 2, pages , Y. Zuo. Projection-based depth functions and associated medians. Ann. Statist., 31: , Y. Zuo and R. Serfling. General notions of statistical depth functions. Ann. Statist., 28-2: , 2000.

23 THANK YOU!