Elements of statistical learning 2nd

Trevor Hastie • Robert Tibshirani • Jerome Friedman The Elements of Statictical Learning This major new edition features many topics not covered in the original, including graphical models, random forests, ensemble methods, least angle regression & path algorithms for the lasso, non-negative matrix factorization, and spectral clustering There is also a chapter on methods for “wide” data (p bigger than n), including multiple testing and false discovery rates Trevor Hastie, Robert Tibshirani, and Jerome Friedman are professors of statistics at Stanford University They are prominent researchers in this area: Hastie and Tibshirani developed generalized additive models and wrote a popular book of that title Hastie codeveloped much of the statistical modeling software and environment in R/S-PLUS and invented principal curves and surfaces Tibshirani proposed the lasso and is co-author of the very successful An Introduction to the Bootstrap Friedman is the co-inventor of many datamining tools including CART, MARS, projection pursuit and gradient boosting S TAT I S T I C S  ---- › springer.com The Elements of Statistical Learning During the past decade there has been an explosion in computation and information technology With it have come vast amounts of data in a variety of fields such as medicine, biology, finance, and marketing The challenge of understanding these data has led to the development of new tools in the field of statistics, and spawned new areas such as data mining, machine learning, and bioinformatics Many of these tools have common underpinnings but are often expressed with different terminology This book describes the important ideas in these areas in a common conceptual framework While the approach is statistical, the emphasis is on concepts rather than mathematics Many examples are given, with a liberal use of color graphics It should be a valuable resource for statisticians and anyone interested in data mining in science or industry The book’s coverage is broad, from supervised learning (prediction) to unsupervised learning The many topics include neural networks, support vector machines, classification trees and boosting—the first comprehensive treatment of this topic in any book Hastie • Tibshirani • Friedman Springer Series in Statistics Springer Series in Statistics Trevor Hastie Robert Tibshirani Jerome Friedman The Elements of Statistical Learning Data Mining, Inference, and Prediction Second Edition This is page v Printer: Opaque this To our parents: Valerie and Patrick Hastie Vera and Sami Tibshirani Florence and Harry Friedman and to our families: Samantha, Timothy, and Lynda Charlie, Ryan, Julie, and Cheryl Melanie, Dora, Monika, and Ildiko vi This is page vii Printer: Opaque this Preface to the Second Edition In God we trust, all others bring data –William Edwards Deming (1900-1993)1 We have been gratified by the popularity of the first edition of The Elements of Statistical Learning This, along with the fast pace of research in the statistical learning field, motivated us to update our book with a second edition We have added four new chapters and updated some of the existing chapters Because many readers are familiar with the layout of the first edition, we have tried to change it as little as possible Here is a summary of the main changes: On the Web, this quote has been widely attributed to both Deming and Robert W Hayden; however Professor Hayden told us that he can claim no credit for this quote, and ironically we could find no “data” confirming that Deming actually said this viii Preface to the Second Edition Chapter Introduction Overview of Supervised Learning Linear Methods for Regression Linear Methods for Classification Basis Expansions and Regularization Kernel Smoothing Methods Model Assessment and Selection Model Inference and Averaging Additive Models, Trees, and Related Methods 10 Boosting and Additive Trees 11 Neural Networks 12 Support Vector Machines and Flexible Discriminants 13 Prototype Methods and Nearest-Neighbors 14 Unsupervised Learning 15 16 17 18 Random Forests Ensemble Learning Undirected Graphical Models High-Dimensional Problems What’s new LAR algorithm and generalizations of the lasso Lasso path for logistic regression Additional illustrations of RKHS Strengths and pitfalls of crossvalidation New example from ecology; some material split off to Chapter 16 Bayesian neural nets and the NIPS 2003 challenge Path algorithm for SVM classifier Spectral clustering, kernel PCA, sparse PCA, non-negative matrix factorization archetypal analysis, nonlinear dimension reduction, Google page rank algorithm, a direct approach to ICA New New New New Some further notes: • Our first edition was unfriendly to colorblind readers; in particular, we tended to favor red/green contrasts which are particularly troublesome We have changed the color palette in this edition to a large extent, replacing the above with an orange/blue contrast • We have changed the name of Chapter from “Kernel Methods” to “Kernel Smoothing Methods”, to avoid confusion with the machinelearning kernel method that is discussed in the context of support vector machines (Chapter 11) and more generally in Chapters and 14 • In the first edition, the discussion of error-rate estimation in Chapter was sloppy, as we did not clearly differentiate the notions of conditional error rates (conditional on the training set) and unconditional rates We have fixed this in the new edition Preface to the Second Edition ix • Chapters 15 and 16 follow naturally from Chapter 10, and the chapters are probably best read in that order • In Chapter 17, we have not attempted a comprehensive treatment of graphical models, and discuss only undirected models and some new methods for their estimation Due to a lack of space, we have specifically omitted coverage of directed graphical models • Chapter 18 explores the “p ≫ N ” problem, which is learning in highdimensional feature spaces These problems arise in many areas, including genomic and proteomic studies, and document classification We thank the many readers who have found the (too numerous) errors in the first edition We apologize for those and have done our best to avoid errors in this new edition We thank Mark Segal, Bala Rajaratnam, and Larry Wasserman for comments on some of the new chapters, and many Stanford graduate and post-doctoral students who offered comments, in particular Mohammed AlQuraishi, John Boik, Holger Hoefling, Arian Maleki, Donal McMahon, Saharon Rosset, Babak Shababa, Daniela Witten, Ji Zhu and Hui Zou We thank John Kimmel for his patience in guiding us through this new edition RT dedicates this edition to the memory of Anna McPhee Trevor Hastie Robert Tibshirani Jerome Friedman Stanford, California August 2008 x Preface to the Second Edition This is page xi Printer: Opaque this Preface to the First Edition We are drowning in information and starving for knowledge –Rutherford D Roger The field of Statistics is constantly challenged by the problems that science and industry brings to its door In the early days, these problems often came from agricultural and industrial experiments and were relatively small in scope With the advent of computers and the information age, statistical problems have exploded both in size and complexity Challenges in the areas of data storage, organization and searching have led to the new field of “data mining”; statistical and computational problems in biology and medicine have created “bioinformatics.” Vast amounts of data are being generated in many fields, and the statistician’s job is to make sense of it all: to extract important patterns and trends, and understand “what the data says.” We call this learning from data The challenges in learning from data have led to a revolution in the statistical sciences Since computation plays such a key role, it is not surprising that much of this new development has been done by researchers in other fields such as computer science and engineering The learning problems that we consider can be roughly categorized as either supervised or unsupervised In supervised learning, the goal is to predict the value of an outcome measure based on a number of input measures; in unsupervised learning, there is no outcome measure, and the goal is to describe the associations and patterns among a set of input measures xii Preface to the First Edition This book is our attempt to bring together many of the important new ideas in learning, and explain them in a statistical framework While some mathematical details are needed, we emphasize the methods and their conceptual underpinnings rather than their theoretical properties As a result, we hope that this book will appeal not just to statisticians but also to researchers and practitioners in a wide variety of fields Just as we have learned a great deal from researchers outside of the field of statistics, our statistical viewpoint may help others to better understand different aspects of learning: There is no true interpretation of anything; interpretation is a vehicle in the service of human comprehension The value of interpretation is in enabling others to fruitfully think about an idea –Andreas Buja We would like to acknowledge the contribution of many people to the conception and completion of this book David Andrews, Leo Breiman, Andreas Buja, John Chambers, Bradley Efron, Geoffrey Hinton, Werner Stuetzle, and John Tukey have greatly influenced our careers Balasubramanian Narasimhan gave us advice and help on many computational problems, and maintained an excellent computing environment Shin-Ho Bang helped in the production of a number of the figures Lee Wilkinson gave valuable tips on color production Ilana Belitskaya, Eva Cantoni, Maya Gupta, Michael Jordan, Shanti Gopatam, Radford Neal, Jorge Picazo, Bogdan Popescu, Olivier Renaud, Saharon Rosset, John Storey, Ji Zhu, Mu Zhu, two reviewers and many students read parts of the manuscript and offered helpful suggestions John Kimmel was supportive, patient and helpful at every phase; MaryAnn Brickner and Frank Ganz headed a superb production team at Springer Trevor Hastie would like to thank the statistics department at the University of Cape Town for their hospitality during the final stages of this book We gratefully acknowledge NSF and NIH for their support of this work Finally, we would like to thank our families and our parents for their love and support Trevor Hastie Robert Tibshirani Jerome Friedman Stanford, California May 2001 The quiet statisticians have changed our world; not by discovering new facts or technical developments, but by changing the ways that we reason, experiment and form our opinions –Ian Hacking This is page xiii Printer: Opaque this Contents Preface to the Second Edition vii Preface to the First Edition xi Introduction Overview of Supervised Learning 2.1 Introduction 2.2 Variable Types and Terminology 2.3 Two Simple Approaches to Prediction: Least Squares and Nearest Neighbors 2.3.1 Linear Models and Least Squares 2.3.2 Nearest-Neighbor Methods 2.3.3 From Least Squares to Nearest Neighbors 2.4 Statistical Decision Theory 2.5 Local Methods in High Dimensions 2.6 Statistical Models, Supervised Learning and Function Approximation 2.6.1 A Statistical Model for the Joint Distribution Pr(X, Y ) 2.6.2 Supervised Learning 2.6.3 Function Approximation 2.7 Structured Regression Models 2.7.1 Difficulty of the Problem 11 11 14 16 18 22 28 28 29 29 32 32 9 References 725 Tropp, J (2006) Just relax: convex programming methods for identifying sparse signals in noise, IEEE Transactions on Information Theory 52: 1030–1051 Valiant, L G (1984) A theory of the learnable, Communications of the ACM 27: 1134–1142 van der Merwe, A and Zidek, J (1980) Multivariate regression analysis and canonical variates, The Canadian Journal of Statistics 8: 27–39 Vapnik, V (1996) The Nature of Statistical Learning Theory, Springer, New York Vapnik, V (1998) Statistical Learning Theory, Wiley, New York Vidakovic, B (1999) Statistical Modeling by Wavelets, Wiley, New York von Luxburg, U (2007) A tutorial on spectral clustering, Statistics and Computing 17(4): 395–416 Wahba, G (1980) Spline bases, regularization, and generalized crossvalidation for solving approximation problems with large quantities of noisy data, Proceedings of the International Conference on Approximation theory in Honour of George Lorenz, Academic Press, Austin, Texas, pp 905–912 Wahba, G (1990) Spline Models for Observational Data, SIAM, Philadelphia Wahba, G., Lin, Y and Zhang, H (2000) GACV for support vector machines, in A Smola, P Bartlett, B Sch¨ olkopf and D Schuurmans (eds), Advances in Large Margin Classifiers, MIT Press, Cambridge, MA., pp 297–311 Wainwright, M (2006) Sharp thresholds for noisy and high-dimensional recovery of sparsity using ℓ1 -constrained quadratic programming, Technical report, Department of Statistics, University of California, Berkeley Wainwright, M J., Ravikumar, P and Lafferty, J D (2007) Highdimensional graphical model selection using ℓ1 -regularized logistic regression, in B Sch¨ olkopf, J Platt and T Hoffman (eds), Advances in Neural Information Processing Systems 19, MIT Press, Cambridge, MA, pp 1465–1472 Wasserman, L (2004) All of Statistics: a Concise Course in Statistical Inference, Springer, New York Weisberg, S (1980) Applied Linear Regression, Wiley, New York 726 References Werbos, P (1974) Beyond Regression, PhD thesis, Harvard University Weston, J and Watkins, C (1999) Multiclass support vector machines, in M Verleysen (ed.), Proceedings of ESANN99, D Facto Press, Brussels Whittaker, J (1990) Graphical Models in Applied Multivariate Statistics, Wiley, Chichester Wickerhauser, M (1994) Adapted Wavelet Analysis from Theory to Software, A.K Peters Ltd, Natick, MA Widrow, B and Hoff, M (1960) Adaptive switching circuits, IRE WESCON Convention record, Vol pp 96-104; Reprinted in Andersen and Rosenfeld (1988) Wold, H (1975) Soft modelling by latent variables: the nonlinear iterative partial least squares (NIPALS) approach, Perspectives in Probability and Statistics, In Honor of M S Bartlett, pp 117–144 Wolpert, D (1992) Stacked generalization, Neural Networks 5: 241–259 Wu, T and Lange, K (2007) The MM alternative to EM, unpublished Wu, T and Lange, K (2008) Coordinate descent procedures for lasso penalized regression, Annals of Applied Statistics 2(1): 224–244 Yee, T and Wild, C (1996) Vector generalized additive models, Journal of the Royal Statistical Society, Series B 58: 481–493 Yuan, M and Lin, Y (2007) Model selection and estimation in regression with grouped variables, Journal of the Royal Statistical Society, Series B 68(1): 49–67 Zhang, P (1993) Model selection via multifold cross-validation, Annals of Statistics 21: 299–311 Zhang, T and Yu, B (2005) Boosting with early stopping: convergence and consistency, Annals of Statistics 33: 1538–1579 Zhao, P and Yu, B (2006) On model selection consistency of lasso, Journal of Machine Learning Research 7: 2541–2563 Zhao, P., Rocha, G and Yu, B (2008) The composite absolute penalties for grouped and hierarchichal variable selection, Annals of Statistics (to appear) Zhu, J and Hastie, T (2004) Classification of gene microarrays by penalized logistic regression, Biostatistics 5(2): 427–443 Zhu, J., Zou, H., Rosset, S and Hastie, T (2005) Multiclass adaboost, Unpublished References 727 Zou, H (2006) The adaptive lasso and its oracle properties, Journal of the American Statistical Association 101: 1418–1429 Zou, H and Hastie, T (2005) Regularization and variable selection via the elastic net, Journal of the Royal Statistical Society Series B 67(2): 301–320 Zou, H., Hastie, T and Tibshirani, R (2006) Sparse principal component analysis, Journal of Computational and Graphical Statistics 15(2): 265–28 Zou, H., Hastie, T and Tibshirani, R (2007) On the degrees of freedom of the lasso, Annals of Statistics 35(5): 2173–2192 728 References This is page 729 Printer: Opaque this Author Index Bakin, S 90 Bellman, R E 22 Bibby, J 94, 135, 441, 539, 559, 578, 630, 679 Bickel, P J 89 Bishop, C 38, 233, 414, 623, 645 Breiman, L 85 Brooks, R J 81 B¨ uhlmann, P 91, 635, 642 Bunea, F 91 Candes, E 86, 89, 613 Chen, S S 68, 94 Cherkassky, V 38, 239 Copas, J B 94, 610 Daubechies, I 92 De Mol, C 92 Defrise, M 92 Donoho, D 68, 86, 91, 94, 613 Duda, R 38, 135 Frank, I 81, 82, 94 Freiha, F 3, 49 Friedman, J 38, 81, 82, 85, 92–94, 121, 126, 657, 661, 667 Fu, W 91, 92 Furnival, G 57 Gill, P 96, 421 Greenshtein, E 91 Hansen, R 93 Hart, P 38, 135 Hastie, T 72, 73, 78, 86, 88, 90, 92–94, 97, 98, 121, 126, 609, 610, 614, 657, 661, 662, 667, 693 Hoefling, H 92, 93, 667 Hoerl, A E 64, 94 Hothorn, T 87, 384 Izenman, A 84 Efron, B 73, 86, 90, 94, 97, 98, 609 Johnstone, I 3, 49, 73, 86, 94, 97, 98, 609 Fan, J 92 Kabalin, J 3, 49 730 Author Index Kennard, R 64, 94 Kent, J 94, 135, 441, 539, 559, 578, 630, 679 Knight, K 91 Lafferty, J 90, 304 Lange, K 92 Lawson, C 93 Li, R 92 Lin, Y 90, 304 Liu, H 90, 304 Mardia, K 94, 135, 441, 539, 559, 578, 630, 679 McNeal, J 3, 49 Meinshausen, N 91, 635, 642 Mulier, F 38, 239 Murray, W 96, 421 Osborne, M 76, 94 Park, M Y 94, 126, 661 Presnell, B 76, 94 Ravikumar, P 90, 304 Redwine, E 3, 49 Ripley, B D 38, 131, 135, 136, 234, 308, 310, 400, 414, 415, 455, 468, 480, 481, 641, 645 Ritov, Y 89, 91 Rocha, G 90 Rosset, S 89, 98, 426, 661 Saunders, M 68, 94 Seber, G 94 Stamey, T 3, 49 Stone, M 81 Stork, D 38, 135 Tao, T 89, 613 Taylor, J 88, 94, 610, 614 Tibshirani, R 73, 78, 86, 88, 90, 92–94, 97, 98, 121, 126, 609, 610, 614, 657, 661, 667 Tropp, J 91 Tsybakov, A 89, 91 Turlach, B 76, 94 van der Merwe, A 84 Vapnik, V 38, 102, 132, 135, 171, 257, 438, 455 Wainwright, M 91 Walther, G 88, 94, 610, 614 Wasserman, L 90, 304 Wegkamp, M 91 Weisberg, S 94 Wilson, R 57 Wold, H 94 Wu, T 92 Yang, N 3, 49 Yu, B 90, 91 Yuan, M 90 Zhang, H 90, 304 Zhao, P 90, 91 Zhu, J 89, 98, 426, 661 Zidek, J 84 Zou, H 72, 78, 92, 662, 693 This is page 731 Printer: Opaque this Index L1 regularization, see Lasso Activation function, 392–395 AdaBoost, 337–346 Adaptive lasso, 92 Adaptive methods, 429 Adaptive nearest neighbor methods, 475–478 Adaptive wavelet filtering, 181 Additive model, 295–304 Adjusted response, 297 Affine set, 130 Affine-invariant average, 482, 540 AIC, see Akaike information criterion Akaike information criterion (AIC), 230 Analysis of deviance, 124 Applications abstracts, 672 aorta, 204 bone, 152 California housing, 371–372, 591 countries, 517 demographics, 379–380 document, 532 flow cytometry, 637 galaxy, 201 heart attack, 122, 146, 207 lymphoma, 674 marketing, 488 microarray, 5, 505, 532 nested spheres, 590 New Zealand fish, 375–379 nuclear magnetic resonance, 176 ozone, 201 prostate cancer, 3, 49, 61, 608 protein mass spectrometry, 664 satellite image, 470 skin of the orange, 429–432 spam, 2, 300–304, 313, 320, 328, 352, 593 vowel, 440, 464 waveform, 451 ZIP code, 4, 404, 536–539 Archetypal analysis, 554–557 Association rules, 492–495, 499– 501 732 Index Automatic relevance determination, 411 Automatic selection of smoothing parameters , 156 B-Spline, 186 Back-propagation, 392–397, 408– 409 Backfitting, 297, 391 Backward selection, 58 stepwise selection, 59 Backward pass, 396 Bagging, 282–288, 409, 587 Basis expansions and regularization, 139–189 Basis functions, 141, 186, 189, 321, 328 Batch learning, 397 Baum–Welch algorithm, 272 Bayes classifier, 21 factor, 234 methods, 233–235, 267–272 rate, 21 Bayesian, 409 Bayesian information criterion (BIC), 233 Benjamini–Hochberg method, 688 Best-subset selection, 57, 610 Between class covariance matrix, 114 Bias, 16, 24, 37, 160, 219 Bias-variance decomposition, 24, 37, 219 Bias-variance tradeoff, 37, 219 BIC, see Bayesian Information Criterion Boltzmann machines, 638–648 Bonferroni method, 686 Boosting, 337–386, 409 as lasso regression, 607–609 exponential loss and AdaBoost, 343 gradient boosting, 358 implementations, 360 margin maximization, 613 numerical optimization, 358 partial-dependence plots, 369 regularization path, 607 shrinkage, 364 stochastic gradient boosting, 365 tree size, 361 variable importance, 367 Bootstrap, 249, 261–264, 267, 271– 282, 587 relationship to Bayesian method, 271 relationship to maximum likelihood method, 267 Bottom-up clustering, 520–528 Bump hunting, see Patient rule induction method Bumping, 290–292 C5.0, 624 Canonical variates, 441 CART, see Classification and regression trees Categorical predictors, 10, 310 Censored data, 674 Classical multidimensional scaling, 570 Classification, 22, 101–137, 305– 317, 417–429 Classification and regression trees (CART), 305–317 Clique, 628 Clustering, 501–528 k-means, 509–510 agglomerative, 523–528 hierarchical, 520–528 Codebook, 515 Combinatorial algorithms, 507 Combining models, 288–290 Committee, 289, 587, 605 Comparison of learning methods, 350–352 Complete data, 276 Index Complexity parameter, 37 Computational shortcuts quadratic penalty, 659 Condensing procedure, 480 Conditional likelihood, 31 Confusion matrix, 301 Conjugate gradients, 396 Consensus, 285–286 Convolutional networks, 407 Coordinate descent, 92, 636, 668 COSSO, 304 Cost complexity pruning, 308 Covariance graph, 631 Cp statistic, 230 Cross-entropy, 308–310 Cross-validation, 241–245 Cubic smoothing spline, 151–153 Cubic spline, 151–153 Curse of dimensionality, 22–26 Dantzig selector, 89 Data augmentation, 276 Daubechies symmlet-8 wavelets, 176 De-correlation, 597 Decision boundary, 13–15, 21 Decision trees, 305–317 Decoder, 515, see encoder Decomposable models, 641 Degrees of freedom in an additive model, 302 in ridge regression, 68 of a tree, 336 of smoother matrices, 153–154, 158 Delta rule, 397 Demmler-Reinsch basis for splines, 156 Density estimation, 208–215 Deviance, 124, 309 Diagonal linear discriminant analysis, 651–654 Dimension reduction, 658 for nearest neighbors, 479 Discrete variables, 10, 310–311 733 Discriminant adaptive nearest neighbor classifier, 475–480 analysis, 106–119 coordinates, 108 functions, 109–110 Dissimilarity measure, 503–504 Dummy variables, 10 Early stopping, 398 Effective degrees of freedom, 17, 68, 153–154, 158, 232, 302, 336 Effective number of parameters, 15, 68, 153–154, 158, 232, 302, 336 Eigenvalues of a smoother matrix, 154 Elastic net, 662 EM algorithm, 272–279 as a maximization-maximization procedure, 277 for two component Gaussian mixture, 272 Encoder, 514–515 Ensemble, 616–623 Ensemble learning, 605–624 Entropy, 309 Equivalent kernel, 156 Error rate, 219–230 Error-correcting codes, 606 Estimates of in-sample prediction error, 230 Expectation-maximization algorithm, see EM algorithm Extra-sample error, 228 False discovery rate, 687–690, 692, 693 Feature, extraction, 150 selection, 409, 658, 681–683 Feed-forward neural networks, 392– 408 734 Index Fisher’s linear discriminant, 106– 119, 438 Flexible discriminant analysis, 440– 445 Forward selection, 58 stagewise, 86, 608 stagewise additive modeling, 342 stepwise, 73 Forward pass algorithm, 395 Fourier transform, 168 Frequentist methods, 267 Function approximation, 28–36 Fused lasso, 666 Gap statistic, 519 Gating networks, 329 Gauss-Markov theorem, 51–52 Gauss-Newton method, 391 Gaussian (normal) distribution, 16 Gaussian graphical model, 630 Gaussian mixtures, 273, 463, 492, 509 Gaussian radial basis functions, 212 GBM, see Gradient boosting GBM package, see Gradient boosting GCV, see Generalized cross-validation GEM (generalized EM), 277 Generalization error, 220 performance, 220 Generalized additive model, 295– 304 Generalized association rules, 497– 499 Generalized cross-validation, 244 Generalized linear discriminant analysis, 438 Generalized linear models, 125 Gibbs sampler, 279–280, 641 for mixtures, 280 Gini index, 309 Global Markov property, 628 Gradient Boosting, 359–361 Gradient descent, 358, 395–397 Graph Laplacian, 545 Graphical lasso, 636 Grouped lasso, 90 Haar basis function, 176 Hammersley-Clifford theorem, 629 Hard-thresholding, 653 Hat matrix, 46 Helix, 582 Hessian matrix, 121 Hidden nodes, 641–642 Hidden units, 393–394 Hierarchical clustering, 520–528 Hierarchical mixtures of experts, 329–332 High-dimensional problems, 649 Hints, 96 Hyperplane, see Separating Hyperplane ICA, see Independent components analysis Importance sampling, 617 In-sample prediction error, 230 Incomplete data, 332 Independent components analysis, 557–570 Independent variables, Indicator response matrix, 103 Inference, 261–294 Information Fisher, 266 observed, 274 Information theory, 236, 561 Inner product, 53, 668, 670 Inputs, 10 Instability of trees, 312 Intercept, 11 Invariance manifold, 471 Invariant metric, 471 Inverse wavelet transform, 179 Index IRLS, see Iteratively reweighted least squares Irreducible error, 224 Ising model, 638 ISOMAP, 572 Isometric feature mapping, 572 Iterative proportional scaling, 585 Iteratively reweighted least squares (IRLS), 121 Jensen’s inequality, 293 Join tree, 629 Junction tree, 629 K-means clustering, 460, 509–514 K-medoid clustering, 515–520 K-nearest neighbor classifiers, 463 Karhunen-Loeve transformation (principal components), 66– 67, 79, 534–539 Karush-Kuhn-Tucker conditions, 133, 420 Kernel classification, 670 density classification, 210 density estimation, 208–215 function, 209 logistic regression, 654 principal component, 547–550 string, 668–669 trick, 660 Kernel methods, 167–176, 208–215, 423–438, 659 Knot, 141, 322 Kriging, 171 Kruskal-Shephard scaling, 570 Kullback-Leibler distance, 561 Lagrange multipliers, 293 Landmark, 539 Laplacian, 545 Laplacian distribution, 72 LAR, see Least angle regression Lasso, 68–69, 86–90, 609, 635, 636, 661 735 fused, 666 Latent factor, 674 variable, 678 Learning, Learning rate, 396 Learning vector quantization, 462 Least angle regression, 73–79, 86, 610 Least squares, 11, 32 Leave-one-out cross-validation, 243 LeNet, 406 Likelihood function, 265, 273 Linear basis expansion, 139–148 Linear combination splits, 312 Linear discriminant function, 106– 119 Linear methods for classification, 101–137 for regression, 43–99 Linear models and least squares, 11 Linear regression of an indicator matrix, 103 Linear separability, 129 Linear smoother, 153 Link function, 296 LLE, see Local linear embedding Local false discovery rate, 693 Local likelihood, 205 Local linear embedding, 572 Local methods in high dimensions, 22–27 Local minima, 400 Local polynomial regression, 197 Local regression, 194, 200 Localization in time/frequency, 175 Loess (local regression), 194, 200 Log-linear model, 639 Log-odds ratio (logit), 119 Logistic (sigmoid) function, 393 Logistic regression, 119–128, 299 Logit (log-odds ratio), 119 Loss function, 18, 21, 219–223, 346 Loss matrix, 310 736 Index Lossless compression, 515 Lossy compression, 515 LVQ, see Learning Vector Quantization Mahalanobis distance, 441 Majority vote, 337 Majorization, 294, 553 Majorize-Minimize algorithm, 294, 584 MAP (maximum aposteriori) estimate, 270 Margin, 134, 418 Market basket analysis, 488, 499 Markov chain Monte Carlo (MCMC) methods, 279 Markov graph, 627 Markov networks, 638–648 MARS, see Multivariate adaptive regression splines MART, see Multiple additive regression trees Maximum likelihood estimation, 31, 261, 265 MCMC, see Markov Chain Monte Carlo Methods MDL, see Minimum description length Mean field approximation, 641 Mean squared error, 24, 285 Memory-based method, 463 Metropolis-Hastings algorithm, 282 Minimum description length (MDL), 235 Minorization, 294, 553 Minorize-Maximize algorithm, 294, 584 Misclassification error, 17, 309 Missing data, 276, 332–333 Missing predictor values, 332–333 Mixing proportions, 214 Mixture discriminant analysis, 449– 455 Mixture modeling, 214–215, 272– 275, 449–455, 692 Mixture of experts, 329–332 Mixtures and the EM algorithm, 272–275 MM algorithm, 294, 584 Mode seekers, 507 Model averaging and stacking, 288 Model combination, 289 Model complexity, 221–222 Model selection, 57, 222–223, 230– 231 Modified regression, 634 Monte Carlo method, 250, 495 Mother wavelet, 178 Multidimensional scaling, 570–572 Multidimensional splines, 162 Multiedit algorithm, 480 Multilayer perceptron, 400, 401 Multinomial distribution, 120 Multiple additive regression trees (MART), 361 Multiple hypothesis testing, 683– 693 Multiple minima, 291, 400 Multiple outcome shrinkage and selection, 84 Multiple outputs, 56, 84, 103–106 Multiple regression from simple univariate regression, 52 Multiresolution analysis, 178 Multivariate adaptive regression splines (MARS), 321–327 Multivariate nonparametric regression, 445 Nadaraya–Watson estimate, 193 Naive Bayes classifier, 108, 210– 211, 694 Natural cubic splines, 144–146 Nearest centroids, 670 Nearest neighbor methods, 463– 483 Nearest shrunken centroids, 651– 654, 694 Network diagram, 392 Neural networks, 389–416 Index Newton’s method (Newton-Raphson procedure), 120–122 Non-negative matrix factorization, 553–554 Nonparametric logistic regression, 299–304 Normal (Gaussian) distribution, 16, 31 Normal equations, 12 Numerical optimization, 395–396 Object dissimilarity, 505–507 Online algorithm, 397 Optimal scoring, 445, 450–451 Optimal separating hyperplane, 132– 135 Optimism of the training error rate, 228–230 Ordered categorical (ordinal) predictor, 10, 504 Ordered features, 666 Orthogonal predictors, 53 Overfitting, 220, 228–230, 364 PageRank, 576 Pairwise distance, 668 Pairwise Markov property, 628 Parametric bootstrap, 264 Partial dependence plots, 369–370 Partial least squares, 80–82, 680 Partition function, 638 Parzen window, 208 Pasting, 318 Path algorithm, 73–79, 86–89, 432 Patient rule induction method(PRIM), 317–321, 499–501 Peeling, 318 Penalization, 607, see regularization Penalized discriminant analysis, 446– 449 Penalized polynomial regression, 171 Penalized regression, 34, 61–69, 171 Penalty matrix, 152, 189 737 Perceptron, 392–416 Piecewise polynomials and splines, 36, 143 Posterior distribution, 268 probability, 233–235, 268 Power method, 577 Pre-conditioning, 681–683 Prediction accuracy, 329 Prediction error, 18 Predictive distribution, 268 PRIM, see Patient rule induction method Principal components, 66–67, 79– 80, 534–539, 547 regression, 79–80 sparse, 550 supervised, 674 Principal curves and surfaces, 541– 544 Principal points, 541 Prior distribution, 268–272 Procrustes average, 540 distance, 539 Projection pursuit, 389–392, 565 regression, 389–392 Prototype classifier, 459–463 Prototype methods, 459–463 Proximity matrices, 503 Pruning, 308 QR decomposition, 55 Quadratic approximations and inference, 124 Quadratic discriminant function, 108, 110 Radial basis function (RBF) network, 392 Radial basis functions, 212–214, 275, 393 Radial kernel, 548 Random forest, 409, 587–604 algorithm, 588 738 Index bias, 596–601 comparison to boosting, 589 example, 589 out-of-bag (oob), 592 overfit, 596 proximity plot, 595 variable importance, 593 variance, 597–601 Rao score test, 125 Rayleigh quotient, 116 Receiver operating characteristic (ROC) curve, 317 Reduced-rank linear discriminant analysis, 113 Regression, 11–14, 43–99, 200–204 Regression spline, 144 Regularization, 34, 167–176 Regularized discriminant analysis, 112–113, 654 Relevance network, 631 Representer of evaluation, 169 Reproducing kernel Hilbert space, 167–176, 428–429 Reproducing property, 169 Responsibilities, 274–275 Ridge regression, 61–68, 650, 659 Risk factor, 122 Robust fitting, 346–350 Rosenblatt’s perceptron learning algorithm, 130 Rug plot, 303 Rulefit, 623 SAM, 690–693, see Significance Analysis of Microarrays Sammon mapping, 571 SCAD, 92 Scaling of the inputs, 398 Schwarz’s criterion, 230–235 Score equations, 120, 265 Self-consistency property, 541–543 Self-organizing map (SOM), 528– 534 Sensitivity of a test, 314–317 Separating hyperplane, 132–135 Separating hyperplanes, 136, 417– 419 Separator, 628 Shape average, 482, 540 Shrinkage methods, 61–69, 652 Sigmoid, 393 Significance Analysis of Microarrays, 690–693 Similarity measure, see Dissimilarity measure Single index model, 390 Singular value decomposition, 64, 535–536, 659 singular values, 535 singular vectors, 535 Sliced inverse regression, 480 Smoother, 139–156, 192–199 matrix, 153 Smoothing parameter, 37, 156–161, 198–199 Smoothing spline, 151–156 Soft clustering, 512 Soft-thresholding, 653 Softmax function, 393 SOM, see Self-organizing map Sparse, 175, 304, 610–613, 636 additive model, 91 graph, 625, 635 Specificity of a test, 314–317 Spectral clustering, 544–547 Spline, 186 additive, 297–299 cubic, 151–153 cubic smoothing, 151–153 interaction, 428 regression, 144 smoothing, 151–156 thin plate, 165 Squared error loss, 18, 24, 37, 219 SRM, see Structural risk minimization Stacking (stacked generalization), 290 Starting values, 397 Statistical decision theory, 18–22 Index Statistical model, 28–29 Steepest descent, 358, 395–397 Stepwise selection, 60 Stochastic approximation, 397 Stochastic search (bumping), 290– 292 Stress function, 570–572 Structural risk minimization (SRM), 239–241 Subset selection, 57–60 Supervised learning, Supervised principal components, 674–681 Support vector classifier, 417–421, 654 multiclass, 657 Support vector machine, 423–437 SURE shrinkage method, 179 Survival analysis, 674 Survival curve, 674 SVD, see Singular value decomposition Symmlet basis, 176 Tangent distance, 471–475 Tanh activation function, 424 Target variables, 10 Tensor product basis, 162 Test error, 220–223 Test set, 220 Thin plate spline, 165 Thinning strategy, 189 Trace of a matrix, 153 Training epoch, 397 Training error, 220–223 Training set, 219–223 Tree for regression, 307–308 Tree-based methods, 305–317 Trees for classification, 308–310 Trellis display, 202 Undirected graph, 625–648 Universal approximator, 390 Unsupervised learning, 2, 485–585 739 Unsupervised learning as supervised learning, 495–497 Validation set, 222 Vapnik-Chervonenkis (VC) dimension, 237–239 Variable importance plot, 594 Variable types and terminology, Variance, 16, 25, 37, 158–161, 219 between, 114 within, 114, 446 Variance reduction, 588 Varying coefficient models, 203– 204 VC dimension, see Vapnik–Chervonenkis dimension Vector quantization, 514–515 Voronoi regions, 510 Wald test, 125 Wavelet basis functions, 176–179 smoothing, 174 transform, 176–179 Weak learner, 383, 605 Weakest link pruning, 308 Webpages, 576 Website for book, Weight decay, 398 Weight elimination, 398 Weights in a neural network, 395 Within class covariance matrix, 114, 446 ... have been gratified by the popularity of the first edition of The Elements of Statistical Learning This, along with the fast pace of research in the statistical learning field, motivated us to update... variety of fields Just as we have learned a great deal from researchers outside of the field of statistics, our statistical viewpoint may help others to better understand different aspects of learning: ... predict the values of the outputs This exercise is called supervised learning We have used the more modern language of machine learning In the statistical literature the inputs are often called the