Linear Support Vector Machine, Linear Support Vector Classification.
is similar to SVC with a linear parameter kernel (SVC(kernel='linear') ), but using liblinear instead of libsvm, so 在选择惩罚和损失函数时更具灵活性,并能更好地扩展到大量样本
SVC(kernel=’linear’)
and LinearSVC()
are similar, except that LinearSVC() is implemented through liblinear; and SVC(kernel='linear') is implemented through libsvm; compared to SVC(kernel ='linear'), LinearSVC) (more flexibility in choosing penalty and loss functions, and scales better to large numbers of samples
1. Algorithm idea
Essentially it is an optimization in SVM, and the principles are similar. For detailed algorithm ideas, you can refer to the blog post: 3. Support Vector Machine Algorithm (SVC, Support Vector Classification) (with supervised learning)
2. Official website API
class sklearn.svm.LinearSVC(penalty='l2', loss='squared_hinge', *, dual='warn', tol=0.0001, C=1.0, multi_class='ovr', fit_intercept=True, intercept_scaling=1, class_weight=None, verbose=0, random_state=None, max_iter=1000)
There are quite a lot of parameters here. For specific parameter usage, you can learn based on the demo provided on the official website and try it more; here are some commonly used parameters for explanation.
Guide package:from sklearn.svm import LinearSVC
①Penalty item penalty
Selection of the penalty term, specifying the specification used in the penalty
The l2 penalty is the standard used by SVC, l1 will cause the coef_vector to be sparse
Regular To put it bluntly, transformation is a constraint on the loss function
In linear regression, L1 regularization, also known as Lasso regression, can produce sparse models
In linear regression, L2 regularization, also known as Ridge regression, can obtain small Parameters to prevent overfitting
specifies the norm used in the penalty. l2 "The penalty is the standard used by SVC. l1" will cause the coef_ vector to be sparse.
'l1': Add L1 regularization
' l2': Add L2 regularization. By default, L2 regularization is the standard for SVC.
Can be selected in SVC()None, but there is no in LinearSVC()
The specific official website details are as follows:
Usage
LinearSVC(penalty='l2')
②Loss function loss
loss, specify the loss function
hinge is the standard SVM loss function (such as used by the SVC class, and squared_hinge is the square of the hinge loss function
The combination of penalty='l1' and loss='hinge' is not supported
Because penalty='l2' is the standard for SVC and loss=' Hinge' is a standard SVM loss function, only such matching ones can be used together
'hinge': standard SVM disfunction number
' squared_hinge': hinge suffix square
The specific official website details are as follows:
Usage
LinearSVC(loss='squared_hinge')
③Regularization parameter C
The strength of the regularization is inversely proportional to C, and the penalty is the square of the L2 regularization. C is a floating point type.
The specific official website details are as follows:
Usage
LinearSVC(C=2.0)
④dual
Whether to choose an algorithm to solve dual or primitive optimization problems, the default is True
"auto": Parameter values will be automatically selected based on the values of n_samples, n_features, loss, multi_class and penalty
If n_samples < n_features, and the optimizer supports the selection of loss, multi_class and penalty, then dual will be set to True, otherwise it will be set to False
The specific official website details are as follows:
⑤Random seed random_state
If you need to control variables for comparison, it is best to set the random seed here to the same integer.
The specific official website details are as follows:
Usage
LinearSVC(random_state=42)
⑤Finally build the model
LinearSVC(penalty=‘l2’,loss=‘squared_hinge’,C=2.0,random_state=42)
3. Code implementation
①Guide package
Here you need to evaluate, train, save and load the model. The following are some necessary packages. If an error is reported during the import process, just install it with pip.
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import joblib
%matplotlib inline
import seaborn as sns
from sklearn.preprocessing import LabelEncoder
from sklearn.model_selection import train_test_split
from sklearn.svm import LinearSVC
from sklearn.metrics import confusion_matrix, classification_report, accuracy_score
②Load the data set
The data set can be simply created by itself in csv format. What I use here is 6 independent variables X and 1 dependent variable Y.
fiber = pd.read_csv("./fiber.csv")
fiber.head(5) #展示下头5条数据信息
③Divide the data set
The first six columns are the independent variable X, and the last column is the dependent variable Y
Official API of commonly used split data set functions:train_test_split
test_size
: Proportion of test set data
train_size
: Proportion of training set data
random_state
: Random seed
shuffle
: Whether to disrupt the data
Because my data set here has a total of 48, training set 0.75, test set 0.25, that is, 36 training sets and 12 test sets
X = fiber.drop(['Grade'], axis=1)
Y = fiber['Grade']
X_train, X_test, y_train, y_test = train_test_split(X,Y,train_size=0.75,test_size=0.25,random_state=42,shuffle=True)
print(X_train.shape) #(36,6)
print(y_train.shape) #(36,)
print(X_test.shape) #(12,6)
print(y_test.shape) #(12,)
④Build LinearSVC model
You can try setting and adjusting the parameters yourself.
lsvc = LinearSVC(penalty='l2',loss='squared_hinge',C=2.0,random_state=42)
⑤Model training
It’s that simple, a fit function can implement model training
lsvc.fit(X_train,y_train)
⑥Model evaluation
Throw the test set in and get the predicted test results
y_pred = lsvc.predict(X_test)
See if the predicted results are consistent with the actual test set results. If consistent, it is 1, otherwise it is 0. The average is the accuracy.
accuracy = np.mean(y_pred==y_test)
print(accuracy)
can also be evaluated by score. The calculation results and ideas are the same. They all look at the probability of the model guessing correctly in all data sets. However, the score function has been encapsulated. Of course, the incoming The parameters are also different, you need to import accuracy_score, from sklearn.metrics import accuracy_score
score = lsvc.score(X_test,y_test)#得分
print(score)
⑦Model testing
Get a piece of data and use the trained model to evaluate
Here are six independent variables. I randomly throw them alltest = np.array([[16,18312.5,6614.5,2842.31,25.23,1147430.19]])
into the model. Get the prediction result, prediction = lsvc.predict(test)
See what the prediction result is and whether it is the same as the correct result, print(prediction)
test = np.array([[16,18312.5,6614.5,2842.31,25.23,1147430.19]])
prediction = lsvc.predict(test)
print(prediction) #[2]
⑧Save the model
lsvc is the model name, which needs to be consistent
The following parameter is the path to save the model
joblib.dump(lsvc, './lsvc.model')#保存模型
⑨Load and use the model
lsvc_yy = joblib.load('./lsvc.model')
test = np.array([[11,99498,5369,9045.27,28.47,3827588.56]])#随便找的一条数据
prediction = lsvc_yy.predict(test)#带入数据,预测一下
print(prediction) #[4]
Complete code
Model training and evaluation does not include ⑧⑨.
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import joblib
%matplotlib inline
import seaborn as sns
from sklearn.preprocessing import LabelEncoder
from sklearn.model_selection import train_test_split
from sklearn.svm import LinearSVC
from sklearn.metrics import confusion_matrix, classification_report, accuracy_score
fiber = pd.read_csv("./fiber.csv")
# 划分自变量和因变量
X = fiber.drop(['Grade'], axis=1)
Y = fiber['Grade']
#划分数据集
X_train, X_test, y_train, y_test = train_test_split(X, Y, random_state=0)
lsvc = LinearSVC(penalty='l2',loss='squared_hinge',C=2.0,random_state=42)
lsvc.fit(X_train,y_train)
y_pred = lsvc.predict(X_test)
accuracy = np.mean(y_pred==y_test)
print(accuracy)
score = lsvc.score(X_test,y_test)#得分
print(score)
test = np.array([[16,18312.5,6614.5,2842.31,25.23,1147430.19]])
prediction = lsvc.predict(test)
print(prediction) #[2]