SVM Kernels Understanding

SVM Kernels Understanding

Open In Colab

%pip install plotly

Collecting plotly

  Downloading plotly-5.20.0-py3-none-any.whl (15.7 MB)

     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 15.7/15.7 MB 451.4 kB/s eta 0:00:0000:0100:01

Requirement already satisfied: tenacity>=6.2.0 in /Users/nipun/mambaforge/lib/python3.10/site-packages (from plotly) (8.2.3)

Requirement already satisfied: packaging in /Users/nipun/mambaforge/lib/python3.10/site-packages (from plotly) (23.1)

Installing collected packages: plotly

Successfully installed plotly-5.20.0

Note: you may need to restart the kernel to use updated packages.

import numpy as np
import matplotlib.pyplot as plt
%matplotlib inline
import plotly.graph_objects as go
from plotly.subplots import make_subplots

%config InlineBackend.figure_format = 'retina'

def linear_kernel(x1, x2):
    return np.dot(x1, x2)

def polynomial_kernel(x, y, p=3):
    return (1 + np.dot(x, y)) ** p

def gaussian_kernel(x, y, gamma=0.5):
    return np.exp(-gamma * np.linalg.norm(x-y)**2)

x_test = np.linspace(-5, 5, 100)
x_point = np.array([1.0])

K_linear_point = np.array([linear_kernel(x_point, xi) for xi in x_test])
K_polynomial_point = np.array([polynomial_kernel(x_point, xi) for xi in x_test])
K_gaussian_point = np.array([gaussian_kernel(x_point, xi) for xi in x_test])
K_gaussian_high_gamma_point = np.array([gaussian_kernel(x_point, xi, gamma=2.0) for xi in x_test])

fig, ax = plt.subplots(nrows=4, sharex=True, figsize=(6, 6))
ax[0].plot(x_test, K_linear_point)
ax[0].set_title('Linear kernel')
ax[1].plot(x_test, K_polynomial_point)
ax[1].set_title('Polynomial kernel')
ax[2].plot(x_test, K_gaussian_point)
ax[2].set_title('Gaussian kernel')
ax[3].plot(x_test, K_gaussian_high_gamma_point)
ax[3].set_title('Gaussian kernel with high $\gamma$')
plt.tight_layout()

for i in range(4):
    ax[i].axvline(x=x_point[0], color='k', linestyle='--', label=r'$x_{sv}$')


plt.legend()

# add ylabel to figure as K(x, x_{sv})
for i in range(4):
    ax[i].set_ylabel(r'$K(x, x_{sv})$')

x = np.linspace(-5, 5, 100)

K_lin = np.array([[linear_kernel(xi, xj) for xi in x] for xj in x])
K_poly = np.array([[polynomial_kernel(xi, xj) for xi in x] for xj in x])
K_gauss = np.array([[gaussian_kernel(xi, xj) for xi in x] for xj in x])

trace_lin = go.Heatmap(z=K_lin, x=x, y=x, colorscale='Viridis', name='Linear Kernel')
fig = go.Figure(data=[trace_lin])
fig.update_layout(title='Linear Kernel', width=500, height=500)
fig.show()

Unable to display output for mime type(s): application/vnd.plotly.v1+json

trace_rbf = go.Heatmap(z=K_gauss, x=x, y=x, colorscale='Viridis', name='Gaussian Kernel')
fig = go.Figure(data=[trace_rbf])
fig.update_layout(title='Gaussian Kernel', width=500, height=500)
fig.show()

Unable to display output for mime type(s): application/vnd.plotly.v1+json

# Plot 1 row 3 cols for various value of gamma in Gaussian kernel using plotly
gammas = [0.1, 1, 5]

fig = make_subplots(rows=1, cols=len(gammas), subplot_titles=[f'Gamma={gamma}' for gamma in gammas])

for i, gamma in enumerate(gammas, 1):
    K_gauss = np.array([[gaussian_kernel(xi, xj, gamma) for xi in x] for xj in x])
    fig.add_trace(go.Heatmap(z=K_gauss, x=x, y=x, colorscale='Viridis'), row=1, col=i)

fig.update_layout(title_text="Gaussian Kernel Heatmaps for Various Gamma Values")
fig.show()

    

Unable to display output for mime type(s): application/vnd.plotly.v1+json

### Understanding how the "similarity" from support vectors to the test point is calculated

from sklearn.datasets import make_blobs
from sklearn.svm import SVC
X, y = make_blobs(n_samples=100, centers=2, random_state=6)
plt.scatter(X[:, 0], X[:, 1], c=y, s=30, cmap=plt.cm.Paired)

clf = SVC(kernel='linear', C = 1e6)
clf.fit(X, y)

SVC(C=1000000.0, kernel='linear')

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

clf.support_vectors_

array([[ 7.27059007, -4.84225716],
       [ 6.49868019, -7.13530714]])

plt.scatter(X[:, 0], X[:, 1], c=y, s=30, cmap=plt.cm.Paired)

# plot scatter with black edge color for support vectors
plt.scatter(clf.support_vectors_[:, 0], clf.support_vectors_[:, 1], s=100,
            linewidth=1, facecolors='none', edgecolors='k')

# test point
test_point = np.array([6, -4])

def decision_function_test(clf, test_point):
    decision_function = 0
    print(f"Test point: {test_point}")
    print("---"*20)
    for i in range(len(clf.support_vectors_)):
        print(f"{i}th support vector: {clf.support_vectors_[i]}")
        print(f'alpha_{i} = {alpha[i]}')
        print(f'y_{clf.support_}[{i}] = {y_plus_minus[clf.support_][i]}')
        print(f'K(x_{i}, x_test) = {linear_kernel(clf.support_vectors_[i], test_point[0])}')
        
        out_i = alpha[i] * y_plus_minus[clf.support_][i] * linear_kernel(clf.support_vectors_[i], test_point[0])
        print(f'alpha_{i} * y_{clf.support_}[{i}] * K(x_{i}, x_test) = {out_i}')
        decision_function += out_i
        print("---"*10)
    print(f"b = {clf.intercept_}")
    print(f"Decision function: {decision_function + clf.intercept_} = {decision_function} + {clf.intercept_}")
    decision_function += clf.intercept_
    return decision_function

clf.dual_coef_

array([[-0.341651,  0.341651]])

test_point = np.array([[7.0, -8.0]])

decision_function_test(clf, test_point)

Test point: [[ 7. -8.]]
------------------------------------------------------------
0th support vector: [ 7.27059007 -4.84225716]
alpha_0 = 0.3416510021250007
y_[48 31][0] = -1
K(x_0, x_test) = 89.63218778443229
alpha_0 * y_[48 31][0] * K(x_0, x_test) = -30.62292677920754
------------------------------
1th support vector: [ 6.49868019 -7.13530714]
alpha_1 = 0.3416510021250007
y_[48 31][1] = 1
K(x_1, x_test) = 102.57321844738127
alpha_1 * y_[48 31][1] * K(x_1, x_test) = 35.04424287373442
------------------------------
b = [-2.87610651]
Decision function: [1.54520958] = 4.42131609452688 + [-2.87610651]

array([1.54520958])

clf.decision_function(test_point)

array([1.54520958])