Least angle regression is a variable selection/shrinkage procedure for high-dimensional data. It is also an algorithm for efficiently finding all knots in the solution path for the aforementioned this regression procedure, as well as for lasso (L1-regularized) linear regression. Fitting the entire solution path is useful for selecting the optimal value of the shrinkage parameter λ for a given dataset, and for the lasso covariance test, which provides the significance of each variable addition along the lasso path.

LARS solution paths are provided by the `lars`

function:

```
lars(X, y; method=:lasso, intercept=true, standardize=true, lambda2=0.0,
use_gram=true, maxiter=typemax(Int), lambda_min=0.0, verbose=false)
```

`X`

is the design matrix and `y`

is the dependent variable. The optional parameters are:

`method`

- either `:lasso`

or `:lars`

.

`intercept`

- whether to fit an intercept in the model. The intercept is
always unpenalized.

`standardize`

- whether to standardize the predictor matrix. In contrast to
linear regression, this affects the algorithm's results. The returned
coefficients are always unstandardized.

`lambda2`

- the elastic net ridge penalty. Zero for pure lasso. Note that the
returned coefficients are the "naive" elastic net coefficients. They can be
adjusted as recommended by Zhou and Hastie (2005) by scaling by `1 + lambda2`

.

`use_gram`

- whether to use a precomputed Gram matrix in computation.

`maxiter`

- maximum number of iterations of the algorithm. If this is
exceeded, an incomplete path is returned. `lambda_min`

- value of λ at which
the algorithm should stop.

`verbose`

- if true, prints information at each step.

The `covtest`

function computes the lasso covariance test based on a LARS path:

`covtest(path, X, y; errorvar)`

`path`

is the output of the LARS function above, and `X`

and `y`

are the
independent and dependent variables used in fitting the path. If specified,
`errorvar`

is the variance of the error. If not specified, the error variance
is computed based on the least squares fit of the full model.

The output of `covtest`

has minor discrepancies with that of the covTest
package. This is
because the covTest package does not take into account the intercept in the
least squares model fit when computing the error variance, which I believe is
incorrect. I have emailed the authors but have yet to receive a response.

LARS.jl is substantially faster than scikit-learn for cases where the number of samples exceeds the number of features, particularly when using a Gram matrix. For cases where the number of features greatly exceeds the number of samples, scikit-learn is still occasionally faster. I am still tracking down the cause.

GLMNet fits the lasso solution path using coordinate descent and supports fitting L1-regularized generalized linear models.

This package is written and maintained by Simon Kornblith simon@simonster.com.

The `lars`

function is derived from code from scikit-learn written by:

- Alexandre Gramfort alexandre.gramfort@inria.fr
- Fabian Pedregosa fabian.pedregosa@inria.fr
- Olivier Grisel olivier.grisel@ensta.org
- Vincent Michel vincent.michel@inria.fr
- Peter Prettenhofer peter.prettenhofer@gmail.com
- Mathieu Blondel mathieu@mblondel.org
- Lars Buitinck L.J.Buitinck@uva.nl

03/21/2014

4 months ago

28 commits