# Trevor Hastie

#### 23 packages on CRAN

#### 1 packages on Bioconductor

Extremely efficient procedures for fitting the entire lasso or elastic-net regularization path for linear regression, logistic and multinomial regression models, Poisson regression, Cox model, multiple-response Gaussian, and the grouped multinomial regression. There are two new and important additions. The family argument can be a GLM family object, which opens the door to any programmed family. This comes with a modest computational cost, so when the built-in families suffice, they should be used instead. The other novelty is the relax option, which refits each of the active sets in the path unpenalized. The algorithm uses cyclical coordinate descent in a path-wise fashion, as described in the papers listed in the URL below.

Efficient procedures for fitting an entire lasso sequence with the cost of a single least squares fit. Least angle regression and infinitesimal forward stagewise regression are related to the lasso, as described in the paper below.

We provide the collection of data-sets used in the book 'An Introduction to Statistical Learning with Applications in R'.

Functions for fitting and working with generalized additive models, as described in chapter 7 of "Statistical Models in S" (Chambers and Hastie (eds), 1991), and "Generalized Additive Models" (Hastie and Tibshirani, 1990).

Mixture and flexible discriminant analysis, multivariate adaptive regression splines (MARS), BRUTO, and vector-response smoothing splines. Hastie, Tibshirani and Friedman (2009) "Elements of Statistical Learning (second edition, chap 12)" Springer, New York.

Iterative methods for matrix completion that use nuclear-norm regularization. There are two main approaches.The one approach uses iterative soft-thresholded svds to impute the missing values. The second approach uses alternating least squares. Both have an "EM" flavor, in that at each iteration the matrix is completed with the current estimate. For large matrices there is a special sparse-matrix class named "Incomplete" that efficiently handles all computations. The package includes procedures for centering and scaling rows, columns or both, and for computing low-rank SVDs on large sparse centered matrices (i.e. principal components)

Computes the entire regularization path for the two-class svm classifier with essentially the same cost as a single SVM fit.

Using overlap grouped-lasso penalties, 'gamsel' selects whether a term in a 'gam' is nonzero, linear, or a non-linear spline (up to a specified max df per variable). It fits the entire regularization path on a grid of values for the overall penalty lambda, both for gaussian and binomial families.

Efficient procedure for fitting regularization paths between L1 and L0, using the MC+ penalty of Zhang, C.H. (2010)<doi:10.1214/09-AOS729>. Implements the methodology described in Mazumder, Friedman and Hastie (2011) <DOI: 10.1198/jasa.2011.tm09738>. Sparsenet computes the regularization surface over both the family parameter and the tuning parameter by coordinate descent.

A direct and flexible method for estimating an ICA model. This approach estimates the densities for each component directly via a tilted gaussian. The tilt functions are estimated via a GAM poisson model. Details can be found in "Elements of Statistical Learning (2nd Edition)" Section 14.7.4

Estimation of a sparse inverse covariance matrix using a lasso (L1) penalty. Facilities are provided for estimates along a path of values for the regularization parameter.

Provides functions for fitting the entire solution path of the Elastic-Net and also provides functions for doing sparse PCA.

L2 penalized logistic regression for both continuous and discrete predictors, with forward stagewise/forward stepwise variable selection procedure.

A path-following algorithm for L1 regularized generalized linear models and Cox proportional hazards model.

Group-Lasso INTERaction-NET. Fits linear pairwise-interaction models that satisfy strong hierarchy: if an interaction coefficient is estimated to be nonzero, then its two associated main effects also have nonzero estimated coefficients. Accommodates categorical variables (factors) with arbitrary numbers of levels, continuous variables, and combinations thereof. Implements the machinery described in the paper "Learning interactions via hierarchical group-lasso regularization" (JCGS 2015, Volume 24, Issue 3). Michael Lim & Trevor Hastie (2015) <DOI:10.1080/10618600.2014.938812>.

Efficient implementations for Sorted L-One Penalized Estimation (SLOPE): generalized linear models regularized with the sorted L1-norm (Bogdan et al. (2015) <doi:10/gfgwzt>). Supported models include ordinary least-squares regression, binomial regression, multinomial regression, and Poisson regression. Both dense and sparse predictor matrices are supported. In addition, the package features predictor screening rules that enable fast and efficient solutions to high-dimensional problems.

Fit a regularized generalized linear model via penalized maximum likelihood. The model is fit for a path of values of the penalty parameter. Fits linear, logistic and Cox models.

Customized training is a simple technique for transductive learning, when the test covariates are known at the time of training. The method identifies a subset of the training set to serve as the training set for each of a few identified subsets in the training set. This package implements customized training for the glmnet() and cv.glmnet() functions.

Fits the Fused Lasso Latent Feature model, which is used for modeling multi-sample aCGH data to identify regions of copy number variation (CNV). Produces a set of features that describe the patterns of CNV and a set of weights that describe the composition of each sample. Also provides functions for choosing the optimal tuning parameters and the appropriate number of features, and for estimating the false discovery rate.

Fit multinomial logistic regression with a penalty on the nuclear norm of the estimated regression coefficient matrix, using proximal gradient descent. Reference: Powers S, Hastie T and Tibshirani R (2018) <doi:10.1177/1471082X18777669>.

Fit a trio model via penalized maximum likelihood. The model is fit for a path of values of the penalty parameter. This package is based on Noah Simon, et al. (2011) <doi:10.1080/10618600.2012.681250>.

Efficient implementations for Sorted L-One Penalized Estimation (SLOPE): generalized linear models regularized with the sorted L1-norm (Bogdan et al. (2015) <doi:10/gfgwzt>) or, equivalently, ordered weighted L1-norm (OWL). Supported models include ordinary least-squares regression, binomial regression, multinomial regression, and Poisson regression. Both dense and sparse predictor matrices are supported. In addition, the package features predictor screening rules that enable fast and efficient solutions to high-dimensional problems.