> For the complete documentation index, see [llms.txt](https://windmising.gitbook.io/liu-yu-bo-play-with-machine-learning/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://windmising.gitbook.io/liu-yu-bo-play-with-machine-learning/src/chapter10-7/10-8.md).

# 10-8 多分类问题中的混淆矩阵

前面小节就针对二分类问题的，这一节把这些指标扩展到多分类问题。

## 加载手写数据集

```python
import numpy as np
from sklearn import datasets

digits = datasets.load_digits()
X = digits.data
y = digits.target

from sklearn.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.8,random_state=666)
```

## 使用逻辑回归训练模型

```python
from sklearn.linear_model import LogisticRegression

log_reg = LogisticRegression()
log_reg.fit(X_train, y_train)
log_reg.score(X_test, y_test)
```

输出：0.93115438108484

## 计算精准率

```python
y_predict = log_reg.predict(X_test)

from sklearn.metrics import precision_score
precision_score(y_test, y_predict)
```

会报错，因为默认只能为二分类问题计算精准率\
解决方法： ![](http://windmissing.github.io/images/2019/201.jpg)

```python
from sklearn.metrics import precision_score
precision_score(y_test, y_predict, average='micro')
```

输出：0.93115438108484

## 计算混淆矩阵并可视化

### 计算混淆矩阵

混淆矩阵天然支持多分类问题。

```python
from sklearn.metrics import confusion_matrix
confusion_matrix(y_test, y_predict)
```

输出： ![](http://windmissing.github.io/images/2019/204.png)

第i行第j列代表真值为i而预测值为j的样本数量。\
对角线位置是预测正确的位置。\
把矩阵可视化，可以直观地看到犯错误比较多的地方。

### 可视化

```python
import matplotlib.pyplot as plt
cfm = confusion_matrix(y_test, y_predict)
plt.matshow(cfm, cmap=plt.cm.gray)   # 映射成灰度值
plt.show()
```

![](http://windmissing.github.io/images/2019/205.png)

越亮的地方代表数值越大。\
图中对角线部分太亮，无法观察到我们关心的部分，改进如下：

```python
row_sums = np.sum(cfm, axis=1)
err_matrix = cfm / row_sums
np.fill_diagonal(err_matrix, 0)
err_matrix
```

![](http://windmissing.github.io/images/2019/206.png)

```python
plt.matshow(err_matrix, cmap=plt.cm.gray)
plt.show()
```

![](http://windmissing.github.io/images/2019/207.png) 图中越亮的地方表示犯错越多的地方。\
针对错误比较多的地方做算法的微调。