![v2-8831fb96dff74c64bdaad9790e1aa61b_1440w.jpg?source=172ae18b](http://img-01.proxy.5ce.com/view/image?&type=2&guid=8ea89a08-4c2f-eb11-8da9-e4434bdf6706&url=https://pic1.zhimg.com/v2-8831fb96dff74c64bdaad9790e1aa61b_1440w.jpg?source=172ae18b)
![v2-38b709c19a7d76bd690fc6f37866a84f_b.jpg](http://img-02.proxy.5ce.com/view/image?&type=2&guid=8ea89a08-4c2f-eb11-8da9-e4434bdf6706&url=https://pic4.zhimg.com/v2-38b709c19a7d76bd690fc6f37866a84f_b.jpg)
今天不单独讲解某个机器学习算法,而是讲解机器学习中常用的工具或者说是方法。一是绘制学习曲线,看模型的好坏程度(过拟合还是欠拟合);而是减少代码量,利用pipeline构造算法流水线。
学习曲线
训练模型通常有三种情况:欠拟合、拟合较好和过拟合。欠拟合一般比较好判别,模型准确度不高都可以说是模型欠拟合。但判断模型是否过拟合,单独看准确度是不可信的,模型越复杂,其准确度越高,也很容易过拟合,这时就需要绘制学习曲线观察模型的拟合情况。
![v2-5be7c68aeedbd444661fd830080f60e2_b.jpg](http://img-01.proxy.5ce.com/view/image?&type=2&guid=8ea89a08-4c2f-eb11-8da9-e4434bdf6706&url=https://pic3.zhimg.com/v2-5be7c68aeedbd444661fd830080f60e2_b.jpg)
绘制流程
- 把数据集划分为多等分(5份或其它)
- 把数据集划分为训练集和测试集
- 以训练集准确性和验证集准确性做为纵坐标,训练集个数作为横坐标。
- 每次增加1等分
绘制函数
在sklearn中,可以通过sklearn.modelselection中的learningcurve来画出学习曲线。这里使用之前KNN算法,通过图可以看出,KNN算法是处于欠拟合状态。
from sklearn.neighbors import
KNeighborsClassifier,RadiusNeighborsClassifier
model1 =
KNeighborsClassifier(n_neighbors=2)
model1.fit(X_train, Y_train)
score1 = model1.score(X_test, Y_test)
from sklearn.model_selection import learning_curve
train_size, train_score, test_score = learning_curve(model1, X, Y, cv=10, train_sizes=np.linspace(0.1,
1.0,
5))
train_scores_mean = np.mean(train_score, axis=1)
train_scores_std = np.std(train_score, axis=1)
test_scores_mean = np.mean(test_score, axis=1)
test_scores_std = np.std(test_score, axis=1)
plt.fill_between(train_size, train_scores_mean - train_scores_std,
train_scores_mean + train_scores_std, alpha=0.1,
color="r")
plt.fill_between(train_size, test_scores_mean - test_scores_std,
test_scores_mean + test_scores_std, alpha=0.1, color="g")
plt.plot(train_size, train_scores_mean,
'o--', color="r",
label="Training score")
plt.plot(train_size, test_scores_mean,
'o-', color="g",
label="Cross-validation score")
plt.grid()
plt.title('Learn Curve for KNN')
plt.legend(loc="best")
![v2-ac1dcccef002f5847a2cbba577b03630_b.jpg](http://img-01.proxy.5ce.com/view/image?&type=2&guid=8ea89a08-4c2f-eb11-8da9-e4434bdf6706&url=https://pic1.zhimg.com/v2-ac1dcccef002f5847a2cbba577b03630_b.jpg)
Pipeline
在之前的线性回归案例中,我们可以加入多项式项来增加模型的精度,但每次都需要先将数据通过PolynomialFeatures转换为新数据,然后再拟合模型,模型预测和评估也需要将测试集进行多项式转换。那能不能将数据处理和模型拟合结合在一起,减少代码量了?答案是可以,通过Pipeline(管道)技术就行。
Pipeline技术
Pipeline 的中间过程由sklearn相适配的转换器(transformer)构成,最后一步是一个estimator(模型)。中间的节点都可以执行fit和transform方法,这样预处理都可以封装进去;最后节点只需要实现fit方法,通常就是我们的模型。流程如下图所示。
![v2-9674b0582b8b37faab47d5a89da3de09_b.jpg](http://img-02.proxy.5ce.com/view/image?&type=2&guid=8ea89a08-4c2f-eb11-8da9-e4434bdf6706&url=https://pic2.zhimg.com/v2-9674b0582b8b37faab47d5a89da3de09_b.jpg)
Pipeline代码
以线性回归为例:
from sklearn.pipeline import
Pipeline
from sklearn.preprocessing import
Normalizer
norm =
Normalizer()
poly =
PolynomialFeatures(2, include_bias=False)
lr =
LinearRegression()
pipeline =
Pipeline([('norm', norm),('poly',poly),('lr', lr)])
pipeline.fit(X_train, y_train)
![v2-d6605b19e8380e6f5a110d3257b4118e_b.jpg](http://img-01.proxy.5ce.com/view/image?&type=2&guid=8ea89a08-4c2f-eb11-8da9-e4434bdf6706&url=https://pic3.zhimg.com/v2-d6605b19e8380e6f5a110d3257b4118e_b.jpg)