python - 我正在尝试使用自定义 RBF 内核在 scikit-learn 中实现 SVM，但它显示错误，该怎么办？

Question

import pandas as pd
import numpy as np
from sklearn import preprocessing, svm
from sklearn.svm import SVC
from sklearn.model_selection import train_test_split
import math
import numpy.linalg as la

df = pd.read_csv("DataWithoutHeader162.csv")
df.columns = ['Temperature','Humidity','Windspeed','Traffic','PM 2.5']
#print(df.head())

forecast_col = 'PM 2.5'
df['label'] = df[forecast_col].shift(1)
df.fillna(value=-99999, inplace=True)

X = np.array(df.drop(['label','PM 2.5'] , 1))
X = preprocessing.scale(X)
df.dropna(inplace = True)

y = np.array(df['label'])
df.dropna(inplace = True)

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.05) 


#kernel definition
def radial_basis(gamma=10):
    return lambda x, y: np.exp(-gamma*la.norm(np.subtract(x, y)))

#SupportVectorMachine with radial_basis Kernel
clf_SVM_radial_basis = SVC(kernel = radial_basis())
clf_SVM_radial_basis.fit(X_train,y_train)
confidence3 = clf_SVM_radial_basis.score(X_test,y_test)
print("Confidence of SVM with radial_basis Kernel = ",(confidence3*100),"%")

此代码显示错误:

Traceback (most recent call last):  
File "F:\MachineLearningPyCodes\SvmOnDelhiAqiDataPrbf.py", line 68, in  
module  
clf_SVM_radial_basis.fit(X_train,y_train)  
File "C:\Python35\lib\site-packages\sklearn\svm\base.py", line 189, in fit  
fit(X, y, sample_weight, solver_type, kernel, random_seed=seed)  
File "C:\Python35\lib\site-packages\sklearn\svm\base.py", line 230, in   
_dense_fit  
if X.shape[0] != X.shape[1]:  
IndexError: tuple index out of range  

我尝试了不同的方法，但我猜我无法按照要求格式化我的数据集，请告诉我一种方法。

Answer 1

我也对 SVC 的 kernel 参数如何工作感到困惑。它不仅仅是您所拥有的内核函数radial_basis；它也是您所拥有的。它实际上必须返回Gram 矩阵。您可以在 sklearn 文档中看到这一点:

http://scikit-learn.org/stable/modules/svm.html#custom-kernels

总结该部分:您有两个选择。

(1) 将 Gram 矩阵直接插入到 fit() 方法中(不仅仅是普通的 X_train)，并使用 kernel='precompulated';或

(2) 编写一个返回 Gram 矩阵的函数，然后您可以将该新函数传递给内核。

这个SO有很好的例子。改编他们写的内容，你可以这样做。我将使用第二种方法，并保留原始的 radial_basis 以供说明。

def radial_basis(x, y, gamma=10):
    return np.exp(-gamma * la.norm(np.subtract(x, y)))

def proxy_kernel(X, Y, K=radial_basis):
    """Another function to return the gram_matrix,
    which is needed in SVC's kernel or fit
    """
    gram_matrix = np.zeros((X.shape[0], Y.shape[0]))
    for i, x in enumerate(X):
        for j, y in enumerate(Y):
            gram_matrix[i, j] = K(x, y)
    return gram_matrix

clf_SVM_radial_basis = SVC(kernel=proxy_kernel) # Note that it's proxy_kernel here now
clf_SVM_radial_basis.fit(X_train, y_train)