Related
I have some trouble plotting the image which is in my head.
I want to visualize the Kernel-trick with Support Vector Machines. So I made some two-dimensional data consisting of two circles (an inner and an outer circle) which should be separated by a hyperplane. Obviously this isn't possible in two dimensions - so I transformed them into 3D. Let n be the number of samples. Now I have an (n,3)-array (3 columns, n rows) X of data points and an (n,1)-array y with labels. Using sklearn I get the linear classifier via
clf = svm.SVC(kernel='linear', C=1000)
clf.fit(X, y)
I already plot the data points as scatter plot via
plt.scatter(X[:, 0], X[:, 1], c=y, s=30, cmap=plt.cm.Paired)
Now I want to plot the separating hyperplane as surface plot. My problem here is the missing explicit representation of the hyperplane because the decision function only yields an implicit hyperplane via decision_function = 0. Therefore I need to plot the level set (of level 0) of an 4-dimensional object.
Since I'm not a python expert I would appreciate if somebody could help me out! And I know that this isn't really the "style" of using a SVM but I need this image as an illustration for my thesis.
Edit: my current "code"
import numpy as np
import matplotlib.pyplot as plt
from sklearn import svm
from sklearn.datasets import make_blobs, make_circles
from tikzplotlib import save as tikz_save
plt.close('all')
# we create 50 separable points
#X, y = make_blobs(n_samples=40, centers=2, random_state=6)
X, y = make_circles(n_samples=50, factor=0.5, random_state=4, noise=.05)
X2, y2 = make_circles(n_samples=50, factor=0.2, random_state=5, noise=.08)
X = np.append(X,X2, axis=0)
y = np.append(y,y2, axis=0)
# shifte X to [0,2]x[0,2]
X = np.array([[item[0] + 1, item[1] + 1] for item in X])
X[X<0] = 0.01
clf = svm.SVC(kernel='rbf', C=1000)
clf.fit(X, y)
plt.scatter(X[:, 0], X[:, 1], c=y, s=30, cmap=plt.cm.Paired)
# plot the decision function
ax = plt.gca()
xlim = ax.get_xlim()
ylim = ax.get_ylim()
# create grid to evaluate model
xx = np.linspace(xlim[0], xlim[1], 30)
yy = np.linspace(ylim[0], ylim[1], 30)
YY, XX = np.meshgrid(yy, xx)
xy = np.vstack([XX.ravel(), YY.ravel()]).T
Z = clf.decision_function(xy).reshape(XX.shape)
# plot decision boundary and margins
ax.contour(XX, YY, Z, colors='k', levels=[-1, 0, 1], alpha=0.5, linestyles=['--','-','--'])
# plot support vectors
ax.scatter(clf.support_vectors_[:, 0], clf.support_vectors_[:, 1], s=100,
linewidth=1, facecolors='none', edgecolors='k')
################## KERNEL TRICK - 3D ##################
trans_X = np.array([[item[0]**2, item[1]**2, np.sqrt(2*item[0]*item[1])] for item in X])
fig = plt.figure()
ax = plt.axes(projection ="3d")
# creating scatter plot
ax.scatter3D(trans_X[:,0],trans_X[:,1],trans_X[:,2], c = y, cmap=plt.cm.Paired)
clf2 = svm.SVC(kernel='linear', C=1000)
clf2.fit(trans_X, y)
ax = plt.gca(projection='3d')
xlim = ax.get_xlim()
ylim = ax.get_ylim()
zlim = ax.get_zlim()
### from here i don't know what to do ###
xx = np.linspace(xlim[0], xlim[1], 3)
yy = np.linspace(ylim[0], ylim[1], 3)
zz = np.linspace(zlim[0], zlim[1], 3)
ZZ, YY, XX = np.meshgrid(zz, yy, xx)
xyz = np.vstack([XX.ravel(), YY.ravel(), ZZ.ravel()]).T
Z = clf2.decision_function(xyz).reshape(XX.shape)
#ax.contour(XX, YY, ZZ, Z, colors='k', levels=[-1, 0, 1], alpha=0.5, linestyles=['--','-','--'])
Desired Output
I want to get something like that.
In general I want to reconstruct what they do in this article, especially "Non-linear transformations".
Part of your question is addressed in this question on linear-kernel SVM. It's a partial answer, because only linear kernels can be represented this way, i.e. thanks to hyperplane coordinates accessible via the estimator when using linear kernel.
Another solution is to find the isosurface with marching_cubes
This solution involves installing the scikit-image toolkit (https://scikit-image.org) which allows to find an isosurface of a given value (here, I considered 0 since it represents the distance to the hyperplane) from the mesh grid of the 3D coordinates.
In the code below (copied from yours), I implement the idea for any kernel (in the example, I used the RBF kernel), and the output is shown beneath the code. Please consider my footnote about 3D plotting with matplotlib, which may be another issue in your case.
import numpy as np
import matplotlib.pyplot as plt
from sklearn import svm
from skimage import measure
from sklearn.datasets import make_blobs, make_circles
from tikzplotlib import save as tikz_save
from mpl_toolkits.mplot3d.art3d import Poly3DCollection
plt.close('all')
# we create 50 separable points
#X, y = make_blobs(n_samples=40, centers=2, random_state=6)
X, y = make_circles(n_samples=50, factor=0.5, random_state=4, noise=.05)
X2, y2 = make_circles(n_samples=50, factor=0.2, random_state=5, noise=.08)
X = np.append(X,X2, axis=0)
y = np.append(y,y2, axis=0)
# shifte X to [0,2]x[0,2]
X = np.array([[item[0] + 1, item[1] + 1] for item in X])
X[X<0] = 0.01
clf = svm.SVC(kernel='rbf', C=1000)
clf.fit(X, y)
plt.scatter(X[:, 0], X[:, 1], c=y, s=30, cmap=plt.cm.Paired)
# plot the decision function
ax = plt.gca()
xlim = ax.get_xlim()
ylim = ax.get_ylim()
# create grid to evaluate model
xx = np.linspace(xlim[0], xlim[1], 30)
yy = np.linspace(ylim[0], ylim[1], 30)
YY, XX = np.meshgrid(yy, xx)
xy = np.vstack([XX.ravel(), YY.ravel()]).T
Z = clf.decision_function(xy).reshape(XX.shape)
# plot decision boundary and margins
ax.contour(XX, YY, Z, colors='k', levels=[-1, 0, 1], alpha=0.5, linestyles=['--','-','--'])
# plot support vectors
ax.scatter(clf.support_vectors_[:, 0], clf.support_vectors_[:, 1], s=100,
linewidth=1, facecolors='none', edgecolors='k')
################## KERNEL TRICK - 3D ##################
trans_X = np.array([[item[0]**2, item[1]**2, np.sqrt(2*item[0]*item[1])] for item in X])
fig = plt.figure()
ax = plt.axes(projection ="3d")
# creating scatter plot
ax.scatter3D(trans_X[:,0],trans_X[:,1],trans_X[:,2], c = y, cmap=plt.cm.Paired)
clf2 = svm.SVC(kernel='rbf', C=1000)
clf2.fit(trans_X, y)
z = lambda x,y: (-clf2.intercept_[0]-clf2.coef_[0][0]*x-clf2.coef_[0][1]*y) / clf2.coef_[0][2]
ax = plt.gca(projection='3d')
xlim = ax.get_xlim()
ylim = ax.get_ylim()
zlim = ax.get_zlim()
### from here i don't know what to do ###
xx = np.linspace(xlim[0], xlim[1], 50)
yy = np.linspace(ylim[0], ylim[1], 50)
zz = np.linspace(zlim[0], zlim[1], 50)
XX ,YY, ZZ = np.meshgrid(xx, yy, zz)
xyz = np.vstack([XX.ravel(), YY.ravel(), ZZ.ravel()]).T
Z = clf2.decision_function(xyz).reshape(XX.shape)
# find isosurface with marching cubes
dx = xx[1] - xx[0]
dy = yy[1] - yy[0]
dz = zz[1] - zz[0]
verts, faces, _, _ = measure.marching_cubes_lewiner(Z, 0, spacing=(1, 1, 1), step_size=2)
verts *= np.array([dx, dy, dz])
verts -= np.array([xlim[0], ylim[0], zlim[0]])
# add as Poly3DCollection
mesh = Poly3DCollection(verts[faces])
mesh.set_facecolor('g')
mesh.set_edgecolor('none')
mesh.set_alpha(0.3)
ax.add_collection3d(mesh)
ax.view_init(20, -45)
plt.savefig('kerneltrick')
Running the code produces the following image with Matplotlib, where the green semi-transparent surface represents the non-linear decision boundary.
Footnote: 3D plotting with matplotlib
Note that Matplotlib 3D is not able to manage the "depth" of objects in some cases, because it can be in conflict with the zorder of this object. This is the reason why sometimes the hyperplane look to be plotted "on top of" the points, even it should be "behind". This issue is a known bug discussed in the matplotlib 3d documentation and in this answer.
If you want to have better rendering results, you may want to use Mayavi, as recommended by the Matplotlib developers, or any other 3D Python plotting library.
I am trying to plot the decision boundaries for the KNeighborsClassifier in Scikit-learn for the Iris dataset. But, the graphs that I am getting don't make much sense to me.
I would expect the boundary between the dark blue and light blue lines go in the direction for the green line that I drew on the picture.
The code that I used to generate it can be found below. It was inspired by Plot the decision boundaries of a VotingClassifier.
What am I missing or not understanding?
# -*- coding: utf-8 -*-
"""
Created on Sat May 30 14:22:05 2020
#author: KamKam
Plotting the decision boundaries for KNearestNeighbours.
"""
# Import required modules.
import matplotlib.pyplot as plt
from sklearn import datasets
from sklearn.neighbors import KNeighborsClassifier
import numpy as np
from matplotlib.colors import ListedColormap
n_neighbors = [1, 3, 9]
# Load the iris dataset.
iris = datasets.load_iris()
X = iris.data[:, 2:4] # Slice features to only contain
y = iris.target
# Set up the data such that it can be inserting into one plot.
# Count the number of each target that are in the dataset.
ylen = y.shape[0]
unique, counts = np.unique(y, return_counts=True)
# Create empty arrays for each of the targets. We only require them to have 2
# features because we are only plotting in 2D.
X0 = np.zeros((counts[0], 2))
X1 = np.zeros((counts[1], 2))
X2 = np.zeros((counts[2], 2))
countX0, countX1, countX2 = 0, 0, 0 #Initialize place holder for interating
# though and adding data to the X arrays.
# Insert data into to newly created arrays.
for i in range(ylen):
if y[i] == 0:
X0[countX0, :] = X[i, :]
countX0 += 1
elif y[i] == 1:
X1[countX1, :] = X[i, :]
countX1 += 1
else:
X2[countX2, :] = X[i, :]
countX2 += 1
h = 0.02 # Step size of the mesh.
plotCount = 0 # Counter for each of the plots that we will be creating.
# Create colour maps.
cmap_light = ListedColormap(['orange', 'cyan', 'cornflowerblue'])
cmap_bold = ListedColormap(['darkorange', 'c', 'darkblue'])
# Initialize plotting. Close all the currently open plots, initialize the
# figure and subplot commands
plt.close('all')
fig, axs = plt.subplots(1, 3)
axs = axs.ravel()
for j in n_neighbors:
# Create the instance od Neighbours classifier and fit the data.
knn = KNeighborsClassifier(n_neighbors=j)
knn.fit(X, y)
# Plot the decision boundary. For that, we will assign a color for each
# point in the mesh [x_min, x_max]x[y_min, y_max]
x_min, x_max = X[:, 0].min() - 1, X[:, 0].max() + 1
y_min, y_max = X[:, 1].min() - 1, X[:, 1].max() + 1
xx, yy = np.meshgrid(np.arange(x_min, x_max, h),
np.arange(y_min, y_max, h))
Z = knn.predict(np.c_[xx.ravel(), yy.ravel()])
# Put the result into a color plot
Z = Z.reshape(xx.shape)
axs[plotCount].pcolormesh(xx, yy, Z, cmap=cmap_bold)
# Plot the training points.
axs[plotCount].scatter(X0[:,0], X0[:,1], c='k', marker='o',
label=iris.target_names[0])
axs[plotCount].scatter(X1[:,0], X1[:,1], c='r', marker='o',
label=iris.target_names[1])
axs[plotCount].scatter(X1[:,0], X2[:,1], c='y', marker='o',
label=iris.target_names[2])
axs[plotCount].set_xlabel('Petal Width')
axs[plotCount].set_ylabel('Petal Length')
axs[plotCount].legend()
axs[plotCount].set_title('n_neighbours = ' + str(j))
plotCount += 1
fig.suptitle('Petal Width vs Length')
plt.show()
The introduction of the arrays X0, X1 and X2 seems to complicate things too much and makes it hard to turn the code more pythonic.
Some things that should be avoided in Python:
the superfluous variable plotCount only to iterate through the axes, can be left out and replaced by for j, ax in zip(n_neighbors, axs).
The contents for X0, X1 and ``X2can be obtained directly viaX[:, 0][y == y_val], X[:, 1][y == y_val]` which moreover allows to write the scatter plots easily in one loop. You can read more about numpy's advanced indexing in this doc.
import matplotlib.pyplot as plt
from sklearn import datasets
from sklearn.neighbors import KNeighborsClassifier
import numpy as np
from matplotlib.colors import ListedColormap
n_neighbors = [1, 3, 9]
# Load the iris dataset.
iris = datasets.load_iris()
X = iris.data[:, 2:4] # Slice features to only contain
y = iris.target
# Set up the data such that it can be inserting into one plot.
# Count the number of each target that are in the dataset.
ylen = y.shape[0]
unique, counts = np.unique(y, return_counts=True)
h = 0.02 # Step size of the mesh.
# Create colour maps.
#cmap_light = ListedColormap(['orange', 'cyan', 'cornflowerblue'])
cmap_bold = ListedColormap(['darkorange', 'c', 'darkblue'])
# Initialize plotting. Close all the currently open plots, initialize the
# figure and subplot commands
plt.close('all')
fig, axs = plt.subplots(1, 3)
axs = axs.ravel()
for j, ax in zip(n_neighbors, axs):
# Create the instance od Neighbours classifier and fit the data.
knn = KNeighborsClassifier(n_neighbors=j)
knn.fit(X, y)
# Plot the decision boundary. For that, we will assign a color for each
# point in the mesh [x_min, x_max]x[y_min, y_max]
x_min, x_max = X[:, 0].min() - h, X[:, 0].max() + h
y_min, y_max = X[:, 1].min() - h, X[:, 1].max() + h
xx, yy = np.meshgrid(np.arange(x_min, x_max, h),
np.arange(y_min, y_max, h))
Z = knn.predict(np.c_[xx.ravel(), yy.ravel()])
# Put the result into a color plot
Z = Z.reshape(xx.shape)
ax.pcolormesh(xx, yy, Z, cmap=cmap_bold)
# Plot the training points.
for y_val, (color, name) in enumerate(zip(['k', 'r', 'y'], iris.target_names)):
ax.scatter(X[:, 0][y == y_val], X[:, 1][y == y_val], c=color, marker='o', label=name)
ax.set_xlabel('Petal Width')
ax.set_ylabel('Petal Length')
ax.legend()
ax.set_title(f'n_neighbours = {j}')
fig.suptitle('Petal Width vs Length')
plt.show()
I'm trying to merge two plots in one:
http://scikit-learn.org/stable/auto_examples/linear_model/plot_sgd_iris.html
http://scikit-learn.org/stable/auto_examples/ensemble/plot_voting_decision_regions.html#sphx-glr-auto-examples-ensemble-plot-voting-decision-regions-py
In the left plot I want to display the decision boundary with the hyperplane corresponding to the OVA classifiers and in the right plot I would like to show the decision probabilities.
This is the code so far:
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sn
from sklearn import datasets
from sklearn import preprocessing
from sklearn.neighbors import KNeighborsClassifier
from sklearn.tree import DecisionTreeClassifier
from sklearn.linear_model import SGDClassifier
from sklearn.svm import SVC
def plot_hyperplane(c, color, fitted_model):
"""
Plot the one-against-all classifiers for the given model.
Parameters
--------------
c : index of the hyperplane to be plot
color : color to be used when drawing the line
fitted_model : the fitted model
"""
xmin, xmax = plt.xlim()
ymin, ymax = plt.ylim()
try:
coef = fitted_model.coef_
intercept = fitted_model.intercept_
except:
return
def line(x0):
return (-(x0 * coef[c, 0]) - intercept[c]) / coef[c, 1]
plt.plot([xmin, xmax], [line(xmin), line(xmax)], ls="--", color=color, zorder=3)
def plot_decision_boundary(X, y, fitted_model, features, targets):
"""
This function plots a model decision boundary as well as it tries to plot
the decision probabilities, if available.
Requires a model fitted with two features only.
Parameters
--------------
X : the data to learn
y : the classification labels
fitted_model : the fitted model
"""
cmap = plt.get_cmap('Set3')
prob = cmap
colors = [cmap(i) for i in np.linspace(0, 1, len(fitted_model.classes_))]
plt.figure(figsize=(9.5, 5))
for i, plot_type in enumerate(['Decision Boundary', 'Decision Probabilities']):
plt.subplot(1, 2, i+1)
mesh_step_size = 0.01 # step size in the mesh
x_min, x_max = X[:, 0].min() - .1, X[:, 0].max() + .1
y_min, y_max = X[:, 1].min() - .1, X[:, 1].max() + .1
xx, yy = np.meshgrid(np.arange(x_min, x_max, mesh_step_size), np.arange(y_min, y_max, mesh_step_size))
# First plot, predicted results using the given model
if i == 0:
Z = fitted_model.predict(np.c_[xx.ravel(), yy.ravel()])
for h, color in zip(fitted_model.classes_, colors):
plot_hyperplane(h, color, fitted_model)
# Second plot, predicted probabilities using the given model
else:
prob = 'RdYlBu_r'
try:
Z = fitted_model.predict_proba(np.c_[xx.ravel(), yy.ravel()])[:, 1]
except:
plt.text(0.4, 0.5, 'Probabilities Unavailable', horizontalalignment='center',
verticalalignment='center', transform=plt.gca().transAxes, fontsize=12)
plt.axis('off')
break
Z = Z.reshape(xx.shape)
# Display Z
plt.imshow(Z, interpolation='nearest', cmap=prob, alpha=0.5,
extent=(x_min, x_max, y_min, y_max), origin='lower', zorder=1)
# Plot the data points
for i, color in zip(fitted_model.classes_, colors):
idx = np.where(y == i)
plt.scatter(X[idx, 0], X[idx, 1], facecolor=color, edgecolor='k', lw=1,
label=iris.target_names[i], cmap=cmap, alpha=0.8, zorder=2)
plt.title(plot_type + '\n' +
str(fitted_model).split('(')[0]+ ' Test Accuracy: ' + str(np.round(fitted_model.score(X, y), 5)))
plt.xlabel(features[0])
plt.ylabel(features[1])
plt.gca().set_aspect('equal')
plt.legend(loc='center left', bbox_to_anchor=(1, 0.5))
plt.tight_layout()
plt.subplots_adjust(top=0.9, bottom=0.08, wspace=0.02)
plt.show()
if __name__ == '__main__':
iris = datasets.load_iris()
X = iris.data[:, [0, 2]]
y = iris.target
scaler = preprocessing.StandardScaler().fit_transform(X)
clf1 = DecisionTreeClassifier(max_depth=4)
clf2 = KNeighborsClassifier(n_neighbors=7)
clf3 = SVC(kernel='rbf', probability=True)
clf4 = SGDClassifier(alpha=0.001, n_iter=100).fit(X, y)
clf1.fit(X, y)
clf2.fit(X, y)
clf3.fit(X, y)
clf4.fit(X, y)
plot_decision_boundary(X, y, clf1, iris.feature_names, iris.target_names[[0, 2]])
plot_decision_boundary(X, y, clf2, iris.feature_names, iris.target_names[[0, 2]])
plot_decision_boundary(X, y, clf3, iris.feature_names, iris.target_names[[0, 2]])
plot_decision_boundary(X, y, clf4, iris.feature_names, iris.target_names[[0, 2]])
And the results:
As can be seen, for the last example (clf4 in the given code) I'm so far unable to plot the hyperplane in the wrong position. I wonder how to correct this. They should be translated to the correct range regarding the used features to fit the model.
Thanks.
Apparently, the problem is the ends of dashed lines representing hyperplanes are not consistent with the final and expected xlim and ylim. A good thing about this case is that you already have x_min, x_max, y_min, y_max defined. So use that and fix xlim and ylim by applying the following 3 lines before plotting hyperplanes (specifically, add in front of your comment line # First plot, predicted results using the given model).
ax = plt.gca()
ax.set_xlim((x_min, x_max), auto=False)
ax.set_ylim((y_min, y_max), auto=False)
in order to visualize the separation of the two classes by a linearSVC, I'm using a plot (defined in the function below)
def show_linearSVC_class_separation(linearSVC: 'LinearSVC', X_test, y_test):
y_decision_score = linearSVC.decision_function(X_test)
# getting the score of the truly positive individuals
y_positive_decision_score = y_decision_score[y_test == 1]
# getting the score of the truly negative individuals
y_negative_decision_score = y_decision_score[y_test == 0]
# counting the distribution of each score value in each class
positive_count = Counter(y_positive_decision_score)
negative_count = Counter(y_negative_decision_score)
# sorting the decision scores to draw a good curve
y_positive_decision_score = np.sort(list(positive_count.keys()))
y_positive_distribution = [positive_count[key] for key in y_positive_decision_score]
y_negative_decision_score = np.sort(list(negative_count.keys()))
y_negative_distribution = [negative_count[key] for key in y_negative_decision_score]
# the alpaha is useful to see the overlaping area between the two classes
plt.fill_between(y_positive_decision_score, 0, y_positive_distribution, color='blue', alpha=0.5, hatch='')
plt.plot(y_positive_decision_score, y_positive_distribution, color='blue', marker='.')
plt.fill_between(y_negative_decision_score, 0, y_negative_distribution, color='red', alpha=0.5, hatch='')
plt.plot(y_negative_decision_score, y_negative_distribution, color='red', marker='.')
plt.legend(['True_positives', 'True_negatives']).draggable()
plt.xlabel('SVM decision_function values')
plt.ylabel('Number of data points')
plt.show()
but, the result is ... pretty ugly, juge by yourself:
I think it's because there is a lot of decision_values that have a counting of one. Maybe an histogramme is the way to go. how can I bucket the decision_values in intervals and count the data points that belong to each interval ?
I need the intervals to have the same length, exemple (length = 1) :
interval || counting
[-7 ; -6] -> 20
]-6 ; -5] -> 30
....
] 5 ; 6] -> 10
Or maybe, there is another way to visualize binary class separation.
to do the visualization, I took inspiration form this blog article Roc curve demonstration.
After some looking around (matplolib and numpy documentation), I finally decided to try to use an histogramme to visualize the class séparation (knowing that I'm working on a multidimensional vector space, ~200k dimensions).
here is the function
''' Plots the seperation plane
Args:
LinearSVC: An LinearSVC instance that was previously fitted (.fit())
'''
def show_linearSVC_class_separation(linearSVC: 'LinearSVC', X_test, y_test):
y_decision_score = linearSVC.decision_function(X_test)
# getting the score of the truly positive individuals
y_positive_decision_score = y_decision_score[y_test == 1]
# getting the score of the truly negative individuals
y_negative_decision_score = y_decision_score[y_test == 0]
# get the (min-1) and the (max +1) scores to be sure to include all the scores in the intervals of the histogramme
_, min_positive = np.modf(y_positive_decision_score.min() - 1)
_, max_positive = np.modf(y_positive_decision_score.max() + 1)
positive_bins = np.arange(min_positive, max_positive + 1)
# get the (min-1) and the (max +1) scores to be sure to include all the scores in the intervals of the histogramme
_, min_negative = np.modf(y_negative_decision_score.min() - 1)
_, max_negative = np.modf(y_negative_decision_score.max() + 1)
negative_bins = np.arange(min_negative, max_negative + 1)
# plot the two histograms, alpha (the transparency) is for the overlapping areas
plt.hist(y_positive_decision_score, bins=positive_bins, alpha=0.5, label='True positives', color='b')
plt.hist(y_negative_decision_score, bins=negative_bins, alpha=0.5, label='True negatives', color='r')
plt.xlabel('SVM decision_function values')
plt.ylabel('Number of data points')
plt.show()
Here is the result for the same example in the question:
Try the following
''' Plots the seperation plane
Args:
LinearSVC: An LinearSVC instance that was previously fitted (.fit())
'''
def show_linearSVC_class_separation(linearSVC: 'LinearSVC', X_test, y_test):
# get the separating hyperplane
w = clf.coef_[0]
a = -w[0] / w[1]
xx = X_test
yy = y_test
# plot the parallels to the separating hyperplane that pass through the
# support vectors
b = clf.support_vectors_[0]
yy_down = a * xx + (b[1] - a * b[0])
b = clf.support_vectors_[-1]
yy_up = a * xx + (b[1] - a * b[0])
# plot the line, the points, and the nearest vectors to the plane
plt.plot(xx, yy, 'k-')
plt.plot(xx, yy_down, 'k--')
plt.plot(xx, yy_up, 'k--')
plt.scatter(clf.support_vectors_[:, 0], clf.support_vectors_[:, 1],
s=80, facecolors='none')
plt.scatter(X[:, 0], X[:, 1], c=Y, cmap=plt.cm.Paired)
plt.axis('tight')
plt.show()
This should generate a similar plot to this:
I could really use a tip to help me plotting a decision boundary to separate to classes of data. I created some sample data (from a Gaussian distribution) via Python NumPy. In this case, every data point is a 2D coordinate, i.e., a 1 column vector consisting of 2 rows. E.g.,
[ 1
2 ]
Let's assume I have 2 classes, class1 and class2, and I created 100 data points for class1 and 100 data points for class2 via the code below (assigned to the variables x1_samples and x2_samples).
mu_vec1 = np.array([0,0])
cov_mat1 = np.array([[2,0],[0,2]])
x1_samples = np.random.multivariate_normal(mu_vec1, cov_mat1, 100)
mu_vec1 = mu_vec1.reshape(1,2).T # to 1-col vector
mu_vec2 = np.array([1,2])
cov_mat2 = np.array([[1,0],[0,1]])
x2_samples = np.random.multivariate_normal(mu_vec2, cov_mat2, 100)
mu_vec2 = mu_vec2.reshape(1,2).T
When I plot the data points for each class, it would look like this:
Now, I came up with an equation for an decision boundary to separate both classes and would like to add it to the plot. However, I am not really sure how I can plot this function:
def decision_boundary(x_vec, mu_vec1, mu_vec2):
g1 = (x_vec-mu_vec1).T.dot((x_vec-mu_vec1))
g2 = 2*( (x_vec-mu_vec2).T.dot((x_vec-mu_vec2)) )
return g1 - g2
I would really appreciate any help!
EDIT:
Intuitively (If I did my math right) I would expect the decision boundary to look somewhat like this red line when I plot the function...
Your question is more complicated than a simple plot : you need to draw the contour which will maximize the inter-class distance. Fortunately it's a well-studied field, particularly for SVM machine learning.
The easiest method is to download the scikit-learn module, which provides a lot of cool methods to draw boundaries: scikit-learn: Support Vector Machines
Code :
# -*- coding: utf-8 -*-
import numpy as np
import matplotlib
from matplotlib import pyplot as plt
import scipy
from sklearn import svm
mu_vec1 = np.array([0,0])
cov_mat1 = np.array([[2,0],[0,2]])
x1_samples = np.random.multivariate_normal(mu_vec1, cov_mat1, 100)
mu_vec1 = mu_vec1.reshape(1,2).T # to 1-col vector
mu_vec2 = np.array([1,2])
cov_mat2 = np.array([[1,0],[0,1]])
x2_samples = np.random.multivariate_normal(mu_vec2, cov_mat2, 100)
mu_vec2 = mu_vec2.reshape(1,2).T
fig = plt.figure()
plt.scatter(x1_samples[:,0],x1_samples[:,1], marker='+')
plt.scatter(x2_samples[:,0],x2_samples[:,1], c= 'green', marker='o')
X = np.concatenate((x1_samples,x2_samples), axis = 0)
Y = np.array([0]*100 + [1]*100)
C = 1.0 # SVM regularization parameter
clf = svm.SVC(kernel = 'linear', gamma=0.7, C=C )
clf.fit(X, Y)
Linear Plot
w = clf.coef_[0]
a = -w[0] / w[1]
xx = np.linspace(-5, 5)
yy = a * xx - (clf.intercept_[0]) / w[1]
plt.plot(xx, yy, 'k-')
MultiLinear Plot
C = 1.0 # SVM regularization parameter
clf = svm.SVC(kernel = 'rbf', gamma=0.7, C=C )
clf.fit(X, Y)
h = .02 # step size in the mesh
# create a mesh to plot in
x_min, x_max = X[:, 0].min() - 1, X[:, 0].max() + 1
y_min, y_max = X[:, 1].min() - 1, X[:, 1].max() + 1
xx, yy = np.meshgrid(np.arange(x_min, x_max, h),
np.arange(y_min, y_max, h))
# Plot the decision boundary. For that, we will assign a color to each
# point in the mesh [x_min, m_max]x[y_min, y_max].
Z = clf.predict(np.c_[xx.ravel(), yy.ravel()])
# Put the result into a color plot
Z = Z.reshape(xx.shape)
plt.contour(xx, yy, Z, cmap=plt.cm.Paired)
Implementation
If you want to implement it yourself, you need to solve the following quadratic equation:
The Wikipedia article
Unfortunately, for non-linear boundaries like the one you draw, it's a difficult problem relying on a kernel trick but there isn't a clear cut solution.
Based on the way you've written decision_boundary you'll want to use the contour function, as Joe noted above. If you just want the boundary line, you can draw a single contour at the 0 level:
f, ax = plt.subplots(figsize=(7, 7))
c1, c2 = "#3366AA", "#AA3333"
ax.scatter(*x1_samples.T, c=c1, s=40)
ax.scatter(*x2_samples.T, c=c2, marker="D", s=40)
x_vec = np.linspace(*ax.get_xlim())
ax.contour(x_vec, x_vec,
decision_boundary(x_vec, mu_vec1, mu_vec2),
levels=[0], cmap="Greys_r")
Which makes:
Those were some great suggestions, thanks a lot for your help! I ended up solving the equation analytically and this is the solution I ended up with (I just want to post it for future reference:
# 2-category classification with random 2D-sample data
# from a multivariate normal distribution
import numpy as np
from matplotlib import pyplot as plt
def decision_boundary(x_1):
""" Calculates the x_2 value for plotting the decision boundary."""
return 4 - np.sqrt(-x_1**2 + 4*x_1 + 6 + np.log(16))
# Generating a Gaussion dataset:
# creating random vectors from the multivariate normal distribution
# given mean and covariance
mu_vec1 = np.array([0,0])
cov_mat1 = np.array([[2,0],[0,2]])
x1_samples = np.random.multivariate_normal(mu_vec1, cov_mat1, 100)
mu_vec1 = mu_vec1.reshape(1,2).T # to 1-col vector
mu_vec2 = np.array([1,2])
cov_mat2 = np.array([[1,0],[0,1]])
x2_samples = np.random.multivariate_normal(mu_vec2, cov_mat2, 100)
mu_vec2 = mu_vec2.reshape(1,2).T # to 1-col vector
# Main scatter plot and plot annotation
f, ax = plt.subplots(figsize=(7, 7))
ax.scatter(x1_samples[:,0], x1_samples[:,1], marker='o', color='green', s=40, alpha=0.5)
ax.scatter(x2_samples[:,0], x2_samples[:,1], marker='^', color='blue', s=40, alpha=0.5)
plt.legend(['Class1 (w1)', 'Class2 (w2)'], loc='upper right')
plt.title('Densities of 2 classes with 25 bivariate random patterns each')
plt.ylabel('x2')
plt.xlabel('x1')
ftext = 'p(x|w1) ~ N(mu1=(0,0)^t, cov1=I)\np(x|w2) ~ N(mu2=(1,1)^t, cov2=I)'
plt.figtext(.15,.8, ftext, fontsize=11, ha='left')
# Adding decision boundary to plot
x_1 = np.arange(-5, 5, 0.1)
bound = decision_boundary(x_1)
plt.plot(x_1, bound, 'r--', lw=3)
x_vec = np.linspace(*ax.get_xlim())
x_1 = np.arange(0, 100, 0.05)
plt.show()
And the code can be found here
EDIT:
I also have a convenience function for plotting decision regions for classifiers that implement a fit and predict method, e.g., the classifiers in scikit-learn, which is useful if the solution cannot be found analytically. A more detailed description how it works can be found here.
You can create your own equation for the boundary:
where you have to find the positions x0 and y0, as well as the constants ai and bi for the radius equation. So, you have 2*(n+1)+2 variables. Using scipy.optimize.leastsq is straightforward for this type of problem.
The code attached below builds the residual for the leastsq penalizing the points outsize the boundary. The result for your problem, obtained with:
x, y = find_boundary(x2_samples[:,0], x2_samples[:,1], n)
ax.plot(x, y, '-k', lw=2.)
x, y = find_boundary(x1_samples[:,0], x1_samples[:,1], n)
ax.plot(x, y, '--k', lw=2.)
using n=1:
using n=2:
usng n=5:
using n=7:
import numpy as np
from numpy import sin, cos, pi
from scipy.optimize import leastsq
def find_boundary(x, y, n, plot_pts=1000):
def sines(theta):
ans = np.array([sin(i*theta) for i in range(n+1)])
return ans
def cosines(theta):
ans = np.array([cos(i*theta) for i in range(n+1)])
return ans
def residual(params, x, y):
x0 = params[0]
y0 = params[1]
c = params[2:]
r_pts = ((x-x0)**2 + (y-y0)**2)**0.5
thetas = np.arctan2((y-y0), (x-x0))
m = np.vstack((sines(thetas), cosines(thetas))).T
r_bound = m.dot(c)
delta = r_pts - r_bound
delta[delta>0] *= 10
return delta
# initial guess for x0 and y0
x0 = x.mean()
y0 = y.mean()
params = np.zeros(2 + 2*(n+1))
params[0] = x0
params[1] = y0
params[2:] += 1000
popt, pcov = leastsq(residual, x0=params, args=(x, y),
ftol=1.e-12, xtol=1.e-12)
thetas = np.linspace(0, 2*pi, plot_pts)
m = np.vstack((sines(thetas), cosines(thetas))).T
c = np.array(popt[2:])
r_bound = m.dot(c)
x_bound = popt[0] + r_bound*cos(thetas)
y_bound = popt[1] + r_bound*sin(thetas)
return x_bound, y_bound
I like the mglearn library to draw decision boundaries. Here is one example from the book "Introduction to Machine Learning with Python" by A. Mueller:
fig, axes = plt.subplots(1, 3, figsize=(10, 3))
for n_neighbors, ax in zip([1, 3, 9], axes):
clf = KNeighborsClassifier(n_neighbors=n_neighbors).fit(X, y)
mglearn.plots.plot_2d_separator(clf, X, fill=True, eps=0.5, ax=ax, alpha=.4)
mglearn.discrete_scatter(X[:, 0], X[:, 1], y, ax=ax)
ax.set_title("{} neighbor(s)".format(n_neighbors))
ax.set_xlabel("feature 0")
ax.set_ylabel("feature 1")
axes[0].legend(loc=3)
If you want to use scikit learn, you can write your code like this:
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from sklearn.linear_model import LogisticRegression
# read data
data = pd.read_csv('ex2data1.txt', header=None)
X = data[[0,1]].values
y = data[2]
# use LogisticRegression
log_reg = LogisticRegression()
log_reg.fit(X, y)
# Coefficient of the features in the decision function. (from theta 1 to theta n)
parameters = log_reg.coef_[0]
# Intercept (a.k.a. bias) added to the decision function. (theta 0)
parameter0 = log_reg.intercept_
# Plotting the decision boundary
fig = plt.figure(figsize=(10,7))
x_values = [np.min(X[:, 1] -5 ), np.max(X[:, 1] +5 )]
# calcul y values
y_values = np.dot((-1./parameters[1]), (np.dot(parameters[0],x_values) + parameter0))
colors=['red' if l==0 else 'blue' for l in y]
plt.scatter(X[:, 0], X[:, 1], label='Logistics regression', color=colors)
plt.plot(x_values, y_values, label='Decision Boundary')
plt.show()
see: Building-a-Logistic-Regression-with-Scikit-learn
Just solved a very similar problem with a different approach (root finding) and wanted to post this alternative as answer here for future reference:
def discr_func(x, y, cov_mat, mu_vec):
"""
Calculates the value of the discriminant function for a dx1 dimensional
sample given covariance matrix and mean vector.
Keyword arguments:
x_vec: A dx1 dimensional numpy array representing the sample.
cov_mat: numpy array of the covariance matrix.
mu_vec: dx1 dimensional numpy array of the sample mean.
Returns a float value as result of the discriminant function.
"""
x_vec = np.array([[x],[y]])
W_i = (-1/2) * np.linalg.inv(cov_mat)
assert(W_i.shape[0] > 1 and W_i.shape[1] > 1), 'W_i must be a matrix'
w_i = np.linalg.inv(cov_mat).dot(mu_vec)
assert(w_i.shape[0] > 1 and w_i.shape[1] == 1), 'w_i must be a column vector'
omega_i_p1 = (((-1/2) * (mu_vec).T).dot(np.linalg.inv(cov_mat))).dot(mu_vec)
omega_i_p2 = (-1/2) * np.log(np.linalg.det(cov_mat))
omega_i = omega_i_p1 - omega_i_p2
assert(omega_i.shape == (1, 1)), 'omega_i must be a scalar'
g = ((x_vec.T).dot(W_i)).dot(x_vec) + (w_i.T).dot(x_vec) + omega_i
return float(g)
#g1 = discr_func(x, y, cov_mat=cov_mat1, mu_vec=mu_vec_1)
#g2 = discr_func(x, y, cov_mat=cov_mat2, mu_vec=mu_vec_2)
x_est50 = list(np.arange(-6, 6, 0.1))
y_est50 = []
for i in x_est50:
y_est50.append(scipy.optimize.bisect(lambda y: discr_func(i, y, cov_mat=cov_est_1, mu_vec=mu_est_1) - \
discr_func(i, y, cov_mat=cov_est_2, mu_vec=mu_est_2), -10,10))
y_est50 = [float(i) for i in y_est50]
Here is the result:
(blue the quadratic case, red the linear case (equal variances)
I know this question has been answered in a very thorough way analytically. I just wanted to share a possible 'hack' to the problem. It is unwieldy but gets the job done.
Start by building a mesh grid of the 2d area and then based on the classifier just build a class map of the entire space. Subsequently detect changes in the decision made row-wise and store the edges points in a list and scatter plot the points.
def disc(x): # returns the class of the point based on location x = [x,y]
temp = 0.5 + 0.5*np.sign(disc0(x)-disc1(x))
# disc0() and disc1() are the discriminant functions of the respective classes
return 0*temp + 1*(1-temp)
num = 200
a = np.linspace(-4,4,num)
b = np.linspace(-6,6,num)
X,Y = np.meshgrid(a,b)
def decColor(x,y):
temp = np.zeros((num,num))
print x.shape, np.size(x,axis=0)
for l in range(num):
for m in range(num):
p = np.array([x[l,m],y[l,m]])
#print p
temp[l,m] = disc(p)
return temp
boundColorMap = decColor(X,Y)
group = 0
boundary = []
for x in range(num):
group = boundColorMap[x,0]
for y in range(num):
if boundColorMap[x,y]!=group:
boundary.append([X[x,y],Y[x,y]])
group = boundColorMap[x,y]
boundary = np.array(boundary)
Sample Decision Boundary for a simple bivariate gaussian classifier
Given two bi-variate normal distributions, you can use Gaussian Discriminant Analysis (GDA) to come up with a decision boundary as the difference between the log of the 2 pdf's.
Here's a way to do it using scipy multivariate_normal (the code is not optimized):
import numpy as np
import matplotlib.pyplot as plt
from scipy.stats import multivariate_normal
from numpy.linalg import norm
from numpy.linalg import inv
from scipy.spatial.distance import mahalanobis
def normal_scatter(mean, cov, p):
size = 100
sigma_x = cov[0,0]
sigma_y = cov[1,1]
mu_x = mean[0]
mu_y = mean[1]
x_ps, y_ps = np.random.multivariate_normal(mean, cov, size).T
x,y = np.mgrid[mu_x-3*sigma_x:mu_x+3*sigma_x:1/size, mu_y-3*sigma_y:mu_y+3*sigma_y:1/size]
grid = np.empty(x.shape + (2,))
grid[:, :, 0] = x; grid[:, :, 1] = y
z = p*multivariate_normal.pdf(grid, mean, cov)
return x_ps, y_ps, x,y,z
# Dist 1
mu_1 = np.array([1, 1])
cov_1 = .5*np.array([[1, 0], [0, 1]])
p_1 = .5
x_ps, y_ps, x,y,z = normal_scatter(mu_1, cov_1, p_1)
plt.plot(x_ps,y_ps,'x')
plt.contour(x, y, z, cmap='Blues', levels=3)
# Dist 2
mu_2 = np.array([2, 1])
#cov_2 = np.array([[2, -1], [-1, 1]])
cov_2 = cov_1
p_2 = .5
x_ps, y_ps, x,y,z = normal_scatter(mu_2, cov_2, p_2)
plt.plot(x_ps,y_ps,'.')
plt.contour(x, y, z, cmap='Oranges', levels=3)
# Decision Boundary
X = np.empty(x.shape + (2,))
X[:, :, 0] = x; X[:, :, 1] = y
g = np.log(p_1*multivariate_normal.pdf(X, mu_1, cov_1)) - np.log(p_2*multivariate_normal.pdf(X, mu_2, cov_2))
plt.contour(x, y, g, [0])
plt.grid()
plt.axhline(y=0, color='k')
plt.axvline(x=0, color='k')
plt.plot([mu_1[0], mu_2[0]], [mu_1[1], mu_2[1]], 'k')
plt.show()
If p_1 != p_2, then you get non-linear boundary. The decision boundary is given by g above.
Then to plot the decision hyper-plane (line in 2D), you need to evaluate g for a 2D mesh, then get the contour which will give a separating line.
You can also assume to have equal co-variance matrices for both distributions, which will give a linear decision boundary. In this case, you can replace the calculation of g in the above code with the following:
W = inv(cov_1).dot(mu_1-mu_2)
x_0 = 1/2*(mu_1+mu_2) - cov_1.dot(np.log(p_1/p_2)).dot((mu_1-mu_2)/mahalanobis(mu_1, mu_2, cov_1))
X = np.empty(x.shape + (2,))
X[:, :, 0] = x; X[:, :, 1] = y
g = (X-x_0).dot(W)
i use this method from this book python-machine-learning-2nd.pdf URL
from matplotlib.colors import ListedColormap
import matplotlib.pyplot as plt
def plot_decision_regions(X, y, classifier, test_idx=None, resolution=0.02):
# setup marker generator and color map
markers = ('s', 'x', 'o', '^', 'v')
colors = ('red', 'blue', 'lightgreen', 'gray', 'cyan')
cmap = ListedColormap(colors[:len(np.unique(y))])
# plot the decision surface
x1_min, x1_max = X[:, 0].min() - 1, X[:, 0].max() + 1
x2_min, x2_max = X[:, 1].min() - 1, X[:, 1].max() + 1
xx1, xx2 = np.meshgrid(np.arange(x1_min, x1_max, resolution),
np.arange(x2_min, x2_max, resolution))
Z = classifier.predict(np.array([xx1.ravel(), xx2.ravel()]).T)
Z = Z.reshape(xx1.shape)
plt.contourf(xx1, xx2, Z, alpha=0.3, cmap=cmap)
plt.xlim(xx1.min(), xx1.max())
plt.ylim(xx2.min(), xx2.max())
for idx, cl in enumerate(np.unique(y)):
plt.scatter(x=X[y == cl, 0],
y=X[y == cl, 1],
alpha=0.8,
c=colors[idx],
marker=markers[idx],
label=cl,
edgecolor='black')
# highlight test samples
if test_idx:
# plot all samples
X_test, y_test = X[test_idx, :], y[test_idx]
plt.scatter(X_test[:, 0],
X_test[:, 1],
c='',
edgecolor='black',
alpha=1.0,
linewidth=1,
marker='o',
s=100,
label='test set')
Since version 1.1, sklearn has a function for this:
https://scikit-learn.org/stable/modules/generated/sklearn.inspection.DecisionBoundaryDisplay.html#sklearn.inspection.DecisionBoundaryDisplay