RIDGE REGRESSION
Performs a PWAS constrained ridge regression on a set of data.
Contents
Function of MOBY-DIC TOOLBOX.
Description
Given a matrix X () and a vector Y (), this routine finds a pwas function f that minimizes the least square residual
The function f is expressed as a weighted sum of alpha basis functions:
In this way, by substituting (2) in (1) the optimization problem can be recast in terms of the weights vector w:
A Tichkonov 0 or 1 order regularization is also performed in order to filter noise. Moreover, constraints on weights w can be imposed in order to reproduce desired behaviours.
At the end, the optimization problem which is solved is the following:
subject to:
The parameter lambda, if not provided, is estimated in order to minimize a cost function. If a test set is provided, lambda is chosen in order to minimize the quadratic error in the test set; if it is not provided a Generalized Cross Validation (GCV) technique is adopted in order to select the best value for lambda (see function GCVcost).
Syntax
[fpwas info] = ridgeRegression(X,Y,P,[Xt],[Yt])
X must be a [ndata x ndim] matrix and Y a [ndata x 1] array. P defines the simplicial partition you want the fpwas function to be defined on. If P is a scalar, each dimension of the domain is subdivided into P intervals. If it is a vector, you specify individually the number of subdivisions per dimension. P can also be a cell array whose i-th element contains the i-th component of the vertices of the simplicial partition (for non-uniform partition). The domain of the pwas function is automatically extrapolated from input data X and Xt. Xt and Yt (optional) represent a test dataset used to estimate the optimal value for the Tikhonov parameter lambda. lambda is chosen in order to minimize | f(Xt) - Yt |^2 being f the pwas function obtained starting from the training dataset (X,Y). If Xt and Yt are not provided, lambda is chosen with a GCV approach. Xt must be a [ndatatest x ndim] matrix and Yt a [ndatatest x 1] array.
fpwas is a pwas object defining the pwas function obtained after the regression.
info is a struct with the following fields:
- wnorm: it is the 2 norm of the weights vector w (for 0 order Tikhonov regularization) or the 2 norm of L w (for 1 order Tikhonov regularization)
- residual: it is the 2 norm of the residual G w - d ( i.e. f(X) - Y )
- test_error: it is the 2 norm of the residual computed in the test set (if provided), i.e. f(Xt) - Yt
[fpwas info] = ridgeRegression(X,Y,P,[Xt],[Yt],D)
As above, but the domain of the pwas function is passed from outside the function. D is a matrix in the form:
[fpwas info] = ridgeRegression(X,Y,P,[Xt],[Yt],options)
options is a structure with the following fields:
- order: Tikhonov regularization order. It can be 0 or 1, default 0.
- lambda: regularization weight. It must be a scalar > 0. If it is not provided it is estimated inside the function.
- nsplits: number of splits of the dataset used to solve the regularised least square problem with an iterative approach (in order to save memory). In practice, instead of solving , you can solve , in which and . nsplits corresponds to n. A low value of nsplits makes ridgeRegression faster but it can result in memroy occupation problems. Default value, 1.
- constraints: array of structures defining optional constraints on the pwas function you want to obtain. Each element of the array defines different constraints. Tre structures have the following fields:
- type: it can be either 'bounds' or 'equality'. If it is 'bounds' constraints in the form lbound <= f(x) <= ubound are imposed. lbound and ubound can be provided through fields lbound and ubound. If they are not provided they are imposed as the minimum and maximum values of the data Y and Yt contained in the dataset. If type is equality, equality constraints in the form f(x) = k can be imposed for x lying on a hyper-plane parallel to the domain components.
- lbound: lower bound for constraints of type 'bounds'
- ubound: upper bound for constraints of type 'bounds'
- variables: needed for equality constraints. It is an array of strings indicating for which set of points you want to impose the constraints. For example if you want to impose a constraint on f, for x1 = 3 and x3 = 2, variables must be ['x1 = 3', 'x3 = 2'].
- value: needed for equality constraints. It is the value you want the function to assume in correspondence of the points defined in field variables. To impose the constraint f(x) = 0, for x2 = 1 and x4 = 5 you gave to set variables = ['x2 = 1', 'x4 = 5'] and value = 0.
- gamma: domain expansion parameter. If the domain D is not provided from outside the function, it is automatically extrapolated from data X (and Xt). The tight domain (Dt) contains exactly data X. Such domain is increased of a factor gamma such as D = gamma Dt. Default value 1.1.
- solver: string specifying the solver you want to use to solve the QP problem (this is necessary only for constrained ridge regression, otherwise the solution is analytical). Possible choices are 'quadprog' (default), 'cvx', 'cplex', 'yalmip' 'clp'.
- verbose: if verbose is set to 1, messages are displayed indicating the status of the ridge regression process.
[fpwas info] = ridgeRegression(X,Y,P,[Xt],[Yt],D,options)
All fields explained above are specified.
Acknowledgements
Contributors:
- Tomaso Poggi (tpoggi@essbilbao.org)
Copyright is with:
- Copyright (C) 2010 University of Genoa, Italy.