python-3.x - Scikit learn/python 中自然文本的有效分类

Question

我希望我的分类算法能够根据一组类别对基于自然语言的原始数据进行分类，当且仅当它满足某个类别的特定阈值准确度(例如 80% 的准确度)，否则我想要我的分类器将该特定原始文本分类为“未分类”类别。我该怎么做？

我的示例数据集:

+----------------------+------------+
| Details              | Category   |
+----------------------+------------+
| Any raw text1        | cat1       |
+----------------------+------------+
| any raw text2        | cat1       |
+----------------------+------------+
| any raw text5        | cat2       |
+----------------------+------------+
| any raw text7        | cat1       |
+----------------------+------------+
| any raw text8        | cat2       |
+----------------------+------------+
| Any raw text4        | cat4       |
+----------------------+------------+
| any raw text5        | cat4       |
+----------------------+------------+
| any raw text6        | cat3       |
+----------------------+------------+

这将是我的训练数据，我将划分相同的数据作为测试集和训练集

import pandas as pd
import numpy as np
import scipy as sp
from sklearn.naive_bayes import MultinomialNB
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score
import matplotlib.pyplot as plt  
from sklearn.model_selection import train_test_split 
data= pd.read_csv('mydata.xls.gold', delimiter='\t',usecols=
['Details','Category'],encoding='utf-8')
target_one=data['Category']
target_list=data['Category'].unique()         
x_train, x_test, y_train, y_test = train_test_split(data.Details, 
data.NUM_CATEGORY, random_state=42)
vect = CountVectorizer(ngram_range=(1,2))
#converting traning features into numeric vector
X_train = vect.fit_transform(x_train.values.astype('U'))
#converting training labels into numeric vector
X_test = vect.transform(x_test.values.astype('U'))
start = time.clock()

mnb = MultinomialNB(alpha =0.13)

mnb.fit(X_train,y_train)

result= mnb.predict(X_test)

print (time.clock()-start)

# mnb.predict_proba(x_test)[0:10,1]
accuracy_score(result,y_test)

我该如何进行？分类器需要设置什么参数吗？ 提前致谢。

Answer 1

您可以使用 predict_proba 结果并使用 columns = target_list 创建 pandas 数据框，然后使用 max和 idxmax找到测试集中每个元素概率最高的类别。完成后，您可以使用 bool 掩码和广播将低于阈值的类别设置为“未分类”

import pandas as pd

df = pd.DataFrame(clf.predict_proba(X_test), columns=target_list)
res_df = pd.DataFrame()

res_df['max_prob'] = df.max(axis=1)
res_df['max_prob_cat'] = df.idxmax(axis=1)

res_df.loc[res_df['max_prob'] < .8, 'max_prob_cat'] = 'unclassified'

df 如下所示

              cat1          cat2          cat3          cat4
0     1.091685e-06  2.257549e-04  9.994661e-01  3.070665e-04
1     2.288312e-02  9.752170e-01  1.783878e-03  1.159706e-04
2     1.980685e-01  3.494765e-01  4.416871e-01  1.076788e-02
3     2.205478e-07  9.999601e-01  3.276864e-05  6.920325e-06
4     2.736805e-03  9.795997e-01  1.718200e-02  4.815429e-04

res_df 看起来像

      max_prob  max_prob_cat
0     0.999466          cat3
1     0.975217          cat2
2     0.441687  unclassified
3     0.999960          cat2
4     0.979600          cat2
5     0.999956          cat2
6     0.998864          cat3
7     0.996888          cat3
8     0.999422          cat1
9     0.994412          cat3
10    0.954508          cat2
11    0.999999          cat2