Important: Please read the installation page for details about how to install the toolboxes. $\newcommand{\dotp}[2]{\langle #1, #2 \rangle}$ $\newcommand{\enscond}[2]{\lbrace #1, #2 \rbrace}$ $\newcommand{\pd}[2]{ \frac{ \partial #1}{\partial #2} }$ $\newcommand{\umin}[1]{\underset{#1}{\min}\;}$ $\newcommand{\umax}[1]{\underset{#1}{\max}\;}$ $\newcommand{\umin}[1]{\underset{#1}{\min}\;}$ $\newcommand{\uargmin}[1]{\underset{#1}{argmin}\;}$ $\newcommand{\norm}[1]{\|#1\|}$ $\newcommand{\abs}[1]{\left|#1\right|}$ $\newcommand{\choice}[1]{ \left\{ \begin{array}{l} #1 \end{array} \right. }$ $\newcommand{\pa}[1]{\left(#1\right)}$ $\newcommand{\diag}[1]{{diag}\left( #1 \right)}$ $\newcommand{\qandq}{\quad\text{and}\quad}$ $\newcommand{\qwhereq}{\quad\text{where}\quad}$ $\newcommand{\qifq}{ \quad \text{if} \quad }$ $\newcommand{\qarrq}{ \quad \Longrightarrow \quad }$ $\newcommand{\ZZ}{\mathbb{Z}}$ $\newcommand{\CC}{\mathbb{C}}$ $\newcommand{\RR}{\mathbb{R}}$ $\newcommand{\EE}{\mathbb{E}}$ $\newcommand{\Zz}{\mathcal{Z}}$ $\newcommand{\Ww}{\mathcal{W}}$ $\newcommand{\Vv}{\mathcal{V}}$ $\newcommand{\Nn}{\mathcal{N}}$ $\newcommand{\NN}{\mathcal{N}}$ $\newcommand{\Hh}{\mathcal{H}}$ $\newcommand{\Bb}{\mathcal{B}}$ $\newcommand{\Ee}{\mathcal{E}}$ $\newcommand{\Cc}{\mathcal{C}}$ $\newcommand{\Gg}{\mathcal{G}}$ $\newcommand{\Ss}{\mathcal{S}}$ $\newcommand{\Pp}{\mathcal{P}}$ $\newcommand{\Ff}{\mathcal{F}}$ $\newcommand{\Xx}{\mathcal{X}}$ $\newcommand{\Mm}{\mathcal{M}}$ $\newcommand{\Ii}{\mathcal{I}}$ $\newcommand{\Dd}{\mathcal{D}}$ $\newcommand{\Ll}{\mathcal{L}}$ $\newcommand{\Tt}{\mathcal{T}}$ $\newcommand{\si}{\sigma}$ $\newcommand{\al}{\alpha}$ $\newcommand{\la}{\lambda}$ $\newcommand{\ga}{\gamma}$ $\newcommand{\Ga}{\Gamma}$ $\newcommand{\La}{\Lambda}$ $\newcommand{\si}{\sigma}$ $\newcommand{\Si}{\Sigma}$ $\newcommand{\be}{\beta}$ $\newcommand{\de}{\delta}$ $\newcommand{\De}{\Delta}$ $\newcommand{\phi}{\varphi}$ $\newcommand{\th}{\theta}$ $\newcommand{\om}{\omega}$ $\newcommand{\Om}{\Omega}$ $\newcommand{\eqdef}{\equiv}$
This tour studies linear regression method in conjunction with regularization. It contrasts ridge regression and the Lasso.
We recommend that after doing this Numerical Tours, you apply it to your own data, for instance using a dataset from <https://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/ LibSVM>.
Disclaimer: these machine learning tours are intended to be overly-simplistic implementations and applications of baseline machine learning methods. For more advanced uses and implementations, we recommend to use a state-of-the-art library, the most well known being <http://scikit-learn.org/ Scikit-Learn>
%matplotlib inline
import numpy as np
import matplotlib.pyplot as plt
%load_ext autoreload
%autoreload 2
Usefull functions to convert to a column/row vectors.
# convert to a column vector
def MakeCol(y): return y.reshape(-1,1)
# convert to a row vector
def MakeRow(y): return y.reshape(1,-1)
# find non zero/true elements
def find(x): return np.nonzero(x)[0]
We test the method on the prostate dataset in $n=97$ samples with features $x_i \in \RR^p$ in dimension $p=8$. The goal is to predict the price value $y_i \in \RR$.
Load the dataset.
from scipy import io
name = 'prostate';
U = io.loadmat('nt_toolbox/data/ml-' + name)
A = U['A']
class_names = U['class_names']
Randomly permute it.
A = A[np.random.permutation(A.shape[0]),:]
Separate the features $X$ from the data $y$ to predict information.
X = A[:,0:-2];
y = MakeCol( A[:,-2] )
c = MakeCol( A[:,-1] )
$n$ is the number of samples, $p$ is the dimensionality of the features,
[n,p] = X.shape
print(n,p)
97 8
Split into training and testing.
I0 = find(c==1) # train
I1 = find(c==0) # test
n0 = I0.size
n1 = n-n0
X0 = X[I0,:]
y0 = y[I0]
X1 = X[I1,:]
y1 = y[I1]
Normalize the features by the mean and std of the training set. This is optional.
mX0 = X0.mean(axis=0)
sX0 = X0.std(axis=0)
X0 = (X0-mX0)/sX0
X1 = (X1-mX0)/sX0
Remove the mean (computed from the test set) to avoid introducing a bias term and a constant regressor. This is optional.
m0 = y0.mean()
y0 = y0-m0
y1 = y1-m0
In order to display in 2-D or 3-D the data, dimensionality is needed. The simplest method is the principal component analysis, which perform an orthogonal linear projection on the principal axsis (eigenvector) of the covariance matrix.
Display the covariance matrix of the training set.
C = X0.transpose().dot(X0)
plt.imshow(C);
u = X0.transpose().dot(y0)
plt.clf
plt.bar(np.arange(1,p+1),u.flatten())
plt.axis('tight');
Compute PCA ortho-basis and the feature in the PCA basis.
U, s, V = np.linalg.svd(X0)
X0r = X0.dot( V.transpose() )
Plot sqrt of the eigenvalues.
plt.plot(s, '.-');
Display the features.
pmax = min(p,8)
k = 0
plt.clf
for i in np.arange(0,pmax):
for j in np.arange(0,pmax):
k = k+1
plt.subplot(pmax,pmax,k)
if i==j:
plt.hist(X0[:,i],6)
plt.axis('tight')
else:
plt.plot(X0[:,j],X0[:,i], '.')
plt.axis('tight')
if i==1:
plt.title(class_names[0][j][0])
plt.tick_params(axis='x', labelbottom=False)
plt.tick_params(axis='y', labelleft=False)
Display the points cloud of feature vectors in 2-D PCA space.
plt.plot(X0r[:,0], X0r[:,1], '.')
plt.axis('equal');
1D plot of the function to regress along the main eigenvector axes.
plt.clf
for i in np.arange(0,3):
plt.subplot(3,1,i+1)
plt.plot(X0r[:,i], y0, '.')
plt.axis('tight')
We look for a linear relationship $ y_i = \dotp{w}{x_i} $ written in matrix format $ y= X w $ where the rows of $X \in \RR^{n \times p}$ stores the features $x_i \in \RR^p$.
Since here $ n > p $, this is an over-determined system, which can solved in the least square sense $$ \umin{ w } \norm{Xw-y}^2 $$ whose solution is given using the Moore-Penrose pseudo-inverse $$ w = (X^\top X)^{-1} X^\top y $$
Least square solution.
w = np.linalg.solve( X0.transpose().dot(X0), X0.transpose().dot(y0) )
Prediction (along 1st eigenvector).
plt.clf
plt.plot( X1.dot(w), '.-' )
plt.plot( y1, '.-' )
plt.axis('tight')
plt.legend(('$y_1$', '$X_1 w$'));
Mean-square error on testing set.
E = np.linalg.norm(X1.dot(w)-y1) / np.linalg.norm(y1)
print(( 'Relative prediction error: ' + str(E) ) );
Relative prediction error: 0.7023447693100116
Although this is not an effective method to solve this problem (a more efficient approach is to use for instance conjugate gradient), one can do a gradient descent to minimize the function $$ \min_w J(w) = \frac{1}{2}\norm{X w-y}^2. $$
def J(w): return 1/2*np.linalg.norm(X0.dot(w)-y0)**2
The gradient of $J$ is $$ \nabla J(w) = X^\top (X w - y). $$
def GradJ(w): return X0.transpose().dot(X0.dot(w)-y0)
The maxium step size allowable by the gradient descent is $$ \tau \leq \tau_\max \eqdef \frac{2}{\norm{XX^\top}_{op}} $$ where $\norm{\cdot}_{op}$ is the maximum singular eigenvalue.
tau = 1/np.linalg.norm(X0,2)**2
Initialize the algorithm to $w=0$.
w = np.zeros((p,1))
One step of gradient descent reads: $$ w \leftarrow w - \tau \nabla J(w). $$
w = w - tau*GradJ(w)
tau_mult = [.1, .5, 1, 1.8]
Exercise 0
Display the evolution of the training error $J(w)$ as a function of the number of iterations. Test for diffetent values of $\tau$.
run -i nt_solutions/ml_2_regression/exo0
The optimal step size to minimize a quadratic function $\dotp{C w}{w}$ is $$ \tau_{\text{opt}} = \frac{2}{\sigma_\min(C) + \sigma_\max(C)} $$
C = X0.transpose().dot(X0)
tau_opt = 2 / ( np.linalg.norm(C,2) + np.linalg.norm(C,-2) )
print(( 'Optimal tau = ' + str( tau_opt * np.linalg.norm(X0,2)**2 ) ) + ' / |X_0 X_0^T|' );
Optimal tau = 1.9032630425437524 / |X_0 X_0^T|
We use SGD (which is not actually a descent algorithm) to minimize the quadratic risk $$ \umin{w} J(w) \eqdef \frac{1}{n} \sum_{i=1}^n f_i(w) = \frac{1}{n} \sum_{i=1}^n \frac{1}{2} ( \dotp{w}{x_i}-y_i )^2 $$ where we used $$ f_i(w) \eqdef \frac{1}{2} ( \dotp{w}{x_i}-y_i )^2$$ The algorithm reads $$ w_{k+1} = w_k - \tau_k \nabla f_{i_k}(w_k)$$ where at each iteration $i_k$ is drawn in $\{1,\ldots,n\}$ uniformly at random.
w = np.zeros((p,1))
Draw $i_k$ are random.
ik = int( np.floor(np.random.rand()*n0) )
Compute $\nabla f_{i_k}(w_k)$.
gk = (X0[ik,:].dot(w)-y0[ik]) * X0[ik,:].transpose()
Set the step size $w_k$ (for convergence is should converge to 0) and perform the update.
tauk = 1/np.linalg.norm(X0,2)**2
w = w - tauk * gk
Test different fixed step size $\tau_k=\tau$.
wopt = np.linalg.solve( X0.transpose().dot(X0), X0.transpose().dot(y0) ) # least square solution
niter = 5000
Jlist = np.zeros((niter,1))
tau_mult = [.05, .3, .8];
for itau in np.arange(0,len(tau_mult)):
tauk = tau_mult[itau]/np.linalg.norm(X0,2)**2
w = np.zeros((p,1))
for i in np.arange(0,niter):
ik = int( np.floor(np.random.rand()*n0) )
gk = (X0[ik,:].dot(w)-y0[ik]) * X0[ik,:].transpose() # stochastic gradient
# gk = X0.transpose().dot( X0.dot(w)-y0 ) # batch gradient
w = w - tauk * gk
Jlist[i] = J(w)
plt.plot(Jlist/J(wopt)-1)
Average on different runs to see the impact of the step size.
niter = 8000
tau_mult = [.05, .8];
nruns = 10 # number of runs to compute the average performance
for itau in np.arange(0,len(tau_mult)):
tauk = tau_mult[itau]/np.linalg.norm(X0,2)**2
Jlist = np.zeros((niter,1))
for iruns in np.arange(0,nruns):
w = np.zeros((p,1))
for i in np.arange(0,niter):
ik = int( np.floor(np.random.rand()*n0) )
gk = (X0[ik,:].dot(w)-y0[ik]) * X0[ik,:].transpose()
w = w - tauk * gk
Jlist[i] = Jlist[i] + J(w)
plt.plot( (Jlist/nruns)/J(wopt)-1 )
Use a decaying step size $\tau_k=\frac{\tau_0}{1+k/k_0}$.
niter = 20000
Jlist = np.zeros((niter,1))
w = np.zeros((p,1))
for i in np.arange(0,niter):
tauk = 1/np.linalg.norm(X0,2)**2 * 1/(1+i/10)
ik = int( np.floor(np.random.rand()*n0) )
gk = (X0[ik,:].dot(w)-y0[ik]) * X0[ik,:].transpose()
w = w - tauk * gk
Jlist[i] = J(w)
plt.plot(Jlist/J(wopt)-1 );
print( Jlist[-1]/J(wopt)-1 )
[14.09087461]
Use a decaying step size $\tau_k=\frac{\tau_0}{1+\sqrt{k/k_0}}$ and average the iteration $\frac{1}{K}\sum_{k<K}w_k$.
niter = 20000
Jlist = np.zeros((niter,1))
w = np.zeros((p,1))
w1 = np.zeros((p,1))
for i in np.arange(0,niter):
tauk = 1/np.linalg.norm(X0,2)**2 * 1/(1+np.sqrt(i/10.0))
ik = int( np.floor(np.random.rand()*n0) )
gk = (X0[ik,:].dot(w)-y0[ik]) * X0[ik,:].transpose()
w = w - tauk * gk
w1 = 1/(i+1) * w + i/(i+1) * w1
Jlist[i] = J(w1)
plt.plot(Jlist/J(wopt)-1 );
Regularization is obtained by introducing a penalty. It is often called ridge regression, and is defined as $$ \umin{ w } \norm{Xw-y}^2 + \lambda \norm{w}^2 $$ where $\lambda>0$ is the regularization parameter.
The solution is given using the following equivalent formula $$ w = (X^\top X + \lambda \text{Id}_p )^{-1} X^\top y, $$ $$ w = X^\top ( XX^\top + \lambda \text{Id}_n)^{-1} y, $$ When $p<n$ (which is the case here), the first formula should be prefered.
In contrast, when the dimensionality $p$ of the feature is very large and there is little data, the second is faster. Furthermore, this second expression is generalizable to Kernel Hilbert space setting, corresponding possibly to $p=+\infty$ for some kernels.
Lambda = .2*np.linalg.norm(X0)**2;
w = np.linalg.solve( X0.transpose().dot(X0) + Lambda*np.eye(p), X0.transpose().dot(y0) )
u = np.linalg.solve( X0.dot(X0.transpose()) + Lambda*np.eye(n0), y0 )
w1 = X0.transpose().dot( u )
print( ('Error (should be 0): ' + str( np.linalg.norm(w-w1)/np.linalg.norm(w) ) ) )
Error (should be 0): 3.856138917165329e-16
Exercise 1
Display the evolution of the test error $E$ as a function of $\lambda$.
run -i nt_solutions/ml_2_regression/exo1
Ridge: 67.90672125664283%
## Insert your code here.
Exercise 2
Display the regularization path, i.e. the evolution of $w$ as a function of $\lambda$.
run -i nt_solutions/ml_2_regression/exo2
## Insert your code here.
In order to perform feature selection (i.e. select a subsect of the features which are the most predictive), one needs to replace the $\ell^2$ regularization penalty by a sparsity inducing regularizer. The most well known is the $\ell^1$ norm $$ \norm{w}_1 \eqdef \sum_i \abs{w_i} . $$
The energy to minimize is $$ \umin{w} J(w) \eqdef \frac{1}{2}\norm{X w-y}^2 + \lambda \norm{w}_1. $$
def J(w,Lambda): return 1/2*np.linalg.norm(X0.dot(w)-y0)**2 + Lambda*np.linalg.norm(w,1)
The simplest iterative algorithm to perform the minimization is the so-called iterative soft thresholding (ISTA), aka proximal gradient aka forward-backward.
It performs first a gradient step (forward) of the smooth part $\frac{1}{2}\norm{X w-y}^2$ of the functional and then a proximal step (backward) step which account for the $\ell^1$ penalty and induce sparsity. This proximal step is the soft-thresholding operator $$ \Ss_s(x) \eqdef \max( \abs{x}-\lambda,0 ) \text{sign}(x). $$
def Soft(x,s): return np.maximum( abs(x)-s, np.zeros(x.shape) ) * np.sign(x)
The ISTA algorithm reads $$ w_{k+1} \eqdef \Ss_{\la\tau}( w_k - \tau X^\top ( X w_k - y ) ), $$ where, to ensure convergence, the step size should verify $ 0 < \tau < 2/\norm{X}^2 $ where $\norm{X}$ is the operator norm.
Display the soft thresholding operator.
t = np.linspace(-5,5,201)
plt.clf
plt.plot(t,Soft(t,2))
plt.axis('tight');
Descent step size.
tau = 1.5/np.linalg.norm(X0,2)**2
Choose a regularization parameter $\la$.
lmax = abs( X0.transpose().dot(y0) ).max()
Lambda = lmax /10
Initialization $w_0$.
w = np.zeros((p,1))
A single ISTA step.
C = X0.transpose().dot(X0)
u = X0.transpose().dot(y0)
def ISTA(w,Lambda,tau): return Soft( w-tau*( C.dot(w)-u ), Lambda*tau )
w = ISTA(w,Lambda,tau)
Exercise 3
Implement the ISTA algorithm, display the convergence of the energy.
run -i nt_solutions/ml_2_regression/exo3
## Insert your code here.
Exercise 4
Compute the test error along the full regularization path. You can start by large $\lambda$ and use a warm restart procedure to reduce the computation time. Compute the classification error. ind optimal lambda isplay error evolution.
run -i nt_solutions/ml_2_regression/exo4
Lasso: 65.42184378732037%
## Insert your code here.
Exercise 5
Display the regularization path, i.e. the evolution of $w$ as a function of $\lambda$. lot(lambda_list, W', 'LineWidth', 2);
run -i nt_solutions/ml_2_regression/exo5
## Insert your code here.
Exercise 6
Compare the optimal weights for ridge and lasso.
run -i nt_solutions/ml_2_regression/exo8