类别变量与文本特征表示

类别变量、独热编码

一组数据:

data = [
    {'name':'张天','born':1996,'died':2077},
    {'name':'李地','born':1986,'died':2067},
    {'name':'吴中','born':1976,'died':2057}
]

其中born和died是数字,对于name要进行文本处理。直观的方式是给每个名字加上一个数字代码,{'张天':1,'李地':2,'吴中':3},这样使名字之间会有大小之分。对于类别变量,往往寻找二值编码,要么是,1,要么不是,0,在机器学习中称独热编码。
上述data的独热编码结果,列名中,对于value是字符串的数据,进行了一些拼接处理:

['born', 'died', 'name=吴中', 'name=张天', 'name=李地']
[[1996 2077    0    1    0]
 [1986 2067    0    0    1]
 [1976 2057    1    0    0]]

利用scikit-learn中的DictVectorizer转换为独热编码,以及系数矩阵的紧凑形式(三元组表示法):

from sklearn.feature_extraction import DictVectorizer
data = [
    {'name':'张天','born':1996,'died':2077},
    {'name':'李地','born':1986,'died':2067},
    {'name':'吴中','born':1976,'died':2057}
]

# sparse=False 稀疏矩阵
vec = DictVectorizer(sparse=False, dtype=int)
# 将矢量化的字典转换成numpy.ndarray。转化后才能用get_feature_names()。
transdata = vec.fit_transform(data)

print(vec.dtype,type(vec))
print(transdata.dtype, type(transdata))

print(vec.get_feature_names())
print(transdata)

# sparse=True 稀疏矩阵的紧凑表示方法
vec2 = DictVectorizer(sparse=True, dtype=int)
transdata2 = vec2.fit_transform(data)
print(vec2.get_feature_names())
print(transdata2)

文本特征

处理文本特征时,较方便的方是用数字计数方式来表示单词或短语

from sklearn.feature_extraction.text import CountVectorizer

# data = ['一个','两个','三个']  中文不太合适
data = ['feature engineering','feature selection','feature extraction feature']

vect = CountVectorizer()
# 计算逆文档频率TF-IDF
Ttransdata = vect.fit_transform(data)
print('Ttransdata:\n',Ttransdata)
print('faeture name:\n',vect.get_feature_names())
print('Ttransdata.toarray:\n',Ttransdata.toarray())

out:

Ttransdata:
   (0, 2)       1
  (0, 0)        1
  (1, 2)        1
  (1, 3)        1
  (2, 2)        2
  (2, 1)        1
faeture name:
 ['engineering', 'extraction', 'feature', 'selection']
Ttransdata.toarray:
 [[1 0 1 0]
 [0 0 1 1]
 [0 1 2 0]]

另一种,根据单词频率计算权重的技术为逆文档频率TF-IDF (term frequency-inverse document frequency)

from sklearn.feature_extraction.text import TfidfVectorizer
# data = ['一个','两个','三个']  中文不太合适
data = ['feature engineering','feature selection','feature extraction feature']

vect = TfidfVectorizer()
# 计算逆文档频率TF-IDF
Ttransdata = vect.fit_transform(data)
print('Ttransdata:\n',Ttransdata)
print('faeture name:\n',vect.get_feature_names())
print('Ttransdata.toarray:\n',Ttransdata.toarray())

out:

Ttransdata:
   (0, 0)       0.8610369959439764
  (0, 2)        0.5085423203783267
  (1, 3)        0.8610369959439764
  (1, 2)        0.5085423203783267
  (2, 1)        0.6461289150464732
  (2, 2)        0.7632282916276542
faeture name:
 ['engineering', 'extraction', 'feature', 'selection']
Ttransdata.toarray:
 [[0.861037   0.         0.50854232 0.        ]
 [0.         0.         0.50854232 0.861037  ]
 [0.         0.64612892 0.76322829 0.        ]]

那么,这些东西有什么用呢?
两个医疗诊断的决策树构建-sklearn、python

posted @ 2021-03-07 14:07  foxerz  阅读(204)  评论(0编辑  收藏  举报