sklearn SVM的参数与R语言的区别

sklearn的SVM函数没有对数据做scale操作，而e1071包的对应函数做数据做了scale。因此在R语言中需要指定scale=FALSE，才会产生跟sklearn类似的结果。

这里以Machine learning with R（机器学习与R语言）一书的letter recognition举例，该数据集也在UCI数据库中，uci letter recognition，这里为了可重复性，使用UCI的数据。

首先在python中，使用pandas读取相应的数据，并将前16000条数据放入训练集，后4000条数据放入测试集，用以评估svm的预测性能。

import pandas as pd
letter_reco_path = "https://archive.ics.uci.edu/ml/machine-learning-databases/letter-recognition/letter-recognition.data"
colnames = [
 "letter", "xbox", "ybox", "width", "height", "onpix", "xbar", "ybar", "x2bar", "y2bar",
            "xybar", "x2ybar", "xy2bar", "xedge", "xedgey", "yedge", "yedgex"   
]
letter_data = pd.read_csv(letter_reco_path, header = None, names = colnames)
training = letter_data.iloc[0:16000,]
testing = letter_data.iloc[16000:, ]
X_train, y_train = training.ix[:, 1:].values, training.ix[:, 0].values
X_train, y_train = training.ix[:, 1:].values, training.ix[:, 0].values

下面使用sklearn的SVC进行SVM的分类，并使用高斯核。

from sklearn.svm import SVC
svm_model = SVC(kernel="rbf", random_state=1071).fit(X_train, y_train)

再对测试集进行预测，得到预测精度0.9722。

svm_pred = svm_model.predict(X_test)
from sklearn.metrics import accuracy_score
accuracy_score(y_test, svm_pred)
0.97224999999999995

同样地，在R语言中，读取UCI对应的数据，把前16000条放入训练集，剩下的放入测试集。

letter_reco_path <- "https://archive.ics.uci.edu/ml/machine-learning-databases/letter-recognition/letter-recognition.data"
colnames <- c("letter", "xbox", "ybox", "width", "height", "onpix", "xbar", "ybar", "x2bar", "y2bar", "xybar", "x2ybar", "xy2bar", "xedge", "xedgey", "yedge", "yedgex")

letter_data <- read.csv( letter_reco_path, header = FALSE,  col.names = colnames)
training_index <- seq.int(1, 16000)

training <- letter_data[training_index, ]
testing <- letter_data[-training_index, ]

通过e1071的svm函数做对应的模型训练，使用高斯核，且对数据不做scale操作，即scale=FALSE。

svm_model2 <- svm(
    letter ~.,
    data = training,
    kernal = "radial",
    type = "C-classification",
    scale = FALSE
)

再通过predict对测试集进行预测，得到精度，0.9725，与sklearn的精度接近。

svm_pred2 <- predict(svm_model2, newdata = testing)
table(svm_pred2 == testing$letter) %>% prop.table

 FALSE   TRUE
0.0275 0.9725

查阅文档，发现sklearn的SVC函数不会对数据做scale操作，而e1071的svm函数默认情况下有scale的操作，需要在实际的使用中注意这种差异。