sklearn支持向量机（SVM）多分类问题

模型

sklearn.svm中的支持向量机：
Classify：SVC、nuSVC、LinearSVC
Regression：SVR、nuSVR、LinearSVR
OneClassSVM

本文采用Classify系列，classify三个模型的区别；参数详解

预处理

import pandas as pd
path = "../Data/classify.csv"
rawdata = pd.read_csv(path) 
X = rawdata.iloc[:,:13]
Y = rawdata.iloc[:,14]  # {”A":0,"B":1,"C":2}
Y = pd.Categorical(Y).codes  # ABC变成123

建模

from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score, recall_score
import sklearn.svm as svm
x_train, x_test, y_train, y_test = \
    train_test_split(X, Y, test_size=0.3)
model = svm.SVC(kernel="linear", decision_function_shape="ovo")

训练

acu_train = model.score(x_train, y_train)
acu_test = model.score(x_test, y_test)
y_pred = model.predict(x_test)
recall = recall_score(y_test, y_pred, average="macro")

多种SVC、核函数对比

def svc(kernel):
    return svm.SVC(kernel=kernel, decision_function_shape="ovo")


def nusvc():
    return svm.NuSVC(decision_function_shape="ovo")


def linearsvc():
    return svm.LinearSVC(multi_class="ovr")


def modelist():
    modelist = []
    kernalist = {"linear", "poly", "rbf", "sigmoid"}
    for each in kernalist:
        modelist.append(svc(each))
    modelist.append(nusvc())
    modelist.append(linearsvc())
    return modelist


def svc_model(model):
    model.fit(x_train, y_train)
    acu_train = model.score(x_train, y_train)
    acu_test = model.score(x_test, y_test)
    y_pred = model.predict(x_test)
    recall = recall_score(y_test, y_pred, average="macro")
    return acu_train, acu_test, recall


def run_svc_model(modelist):
    result = {"kernel": [],
              "acu_train": [],
              "acu_test": [],
              "recall": []
              }

    for model in modelist:
        acu_train, acu_test, recall = svc_model(model)
        try:
            result["kernel"].append(model.kernel)
        except:
            result["kernel"].append(None)
        result["acu_train"].append(acu_train)
        result["acu_test"].append(acu_test)
        result["recall"].append(recall)

    return pd.DataFrame(result)

run_svc_model(modelist())

对比的结果：

优化linear核函数的SVC的惩罚系数

惩罚系数（C=）越高，对错误分类的惩罚越大，模型训练时的准确率就会提高。但若惩罚系数过高，不仅增加了计算资源的消耗，还可能导致模型过拟合，泛化能力减弱。

def test_c():
    result = {"C": [],
              "acu_train": [],
              "acu_test": [],
              "recall": []
              }
    for c in range(10, 101, 10):
        model = svm.SVC(kernel="linear", C=c, decision_function_shape="ovo")
        acu_train, acu_test, recall = svc_model(model)
        result["C"].append(c)
        result["acu_train"].append(acu_train)
        result["acu_test"].append(acu_test)
        result["recall"].append(recall)
    df = pd.DataFrame(result)
    return df
test_c()

结果

大概60的时候是最优的。

其他模型的优化

其他模型也可以通过调节其他参数优化，比如poly为核函数的SVC模型，可以调节多项式的次数进行优化：

def test_degree():
    result = {"degree": [],
              "acu_train": [],
              "acu_test": [],
              "recall": []
              }
    for d in range(1, 11, 1):
        model = svm.SVC(kernel="poly", degree=d, decision_function_shape="ovo")
        acu_train, acu_test, recall = svc_model(model)
        result["degree"].append(d)
        result["acu_train"].append(acu_train)
        result["acu_test"].append(acu_test)
        result["recall"].append(recall)
    df = pd.DataFrame(result)
    return df
    
test_degree()

结果（纵轴是测试集上的准确率）

一般是多项式的次数越高，准确率就越高。但是还是没有达到linear的80%。而且消耗计算资源。